Google Deploys Autonomous Agents and Gemini 3.5 in Race Toward AGI

Google I/O '26 Keynote — Google

Google

Google contends that breakthroughs in inference efficiency and multi-agent orchestration have transformed software from passive prompts into autonomous 24/7 task execution as artificial general intelligence nears.

AI-assisted summary of Google I/O '26 Keynote — original source. Check the original for full context.

Source published:

Key ideas

  • Google marked 10 years as an AI-first company with monthly token processing reaching 3.2 quadrillion tokens, supported by annual capital expenditures climbing toward $180B–$190B and 8th-generation dual-chip TPUs. (1:54)
  • Gemini 3.5 Flash and Google Antigravity 2.0 shifted software engineering toward autonomous multi-agent execution, highlighted by the generation of a functional operating system in 12 hours for under $1,000 in API credits. (27:14)
  • Sundar Pichai unveiled Gemini Spark, a 24/7 cloud-hosted personal agent running on dedicated virtual machines that completes complex, asynchronous workflows across Google Workspace and external services even when devices are closed. (33:35)
  • Google introduced open commerce standards including the Universal Commerce Protocol (co-steered by Amazon, Meta, Microsoft, and Stripe) and the Agent Payments Protocol to enable multi-merchant checkout and autonomous agent transactions. (58:45)
  • Android XR expanded into physical eyewear through lightweight, screenless audio glasses designed with Samsung, Warby Parker, and Gentle Monster, scheduled for commercial release in the fall. (93:00)

Detailed notes

## Background Published by Google on YouTube on May 19, 2026, this video documents the Google I/O '26 Keynote presentation. According to the show notes, the event highlights Google's latest product news, platform announcements, and artificial intelligence advancements across consumer and developer ecosystems.

## TL;DR Commemorating a decade since pivoting to become an AI-first company, Google presents its full-stack AI roadmap anchored by next-generation TPU v8 silicon, Gemini Omni world models, and the hyper-efficient Gemini 3.5 Flash. The keynote showcases a decisive transition into autonomous agents, unveiling Antigravity 2.0 for end-to-end software development, Gemini Spark for 24/7 background task execution, and generative UI alongside the Universal Commerce Protocol in Search. The event also previews consumer hardware innovations with Android XR intelligent eyewear in partnership with Samsung, Gentle Monster, and Warby Parker, before concluding with Demis Hassabis outlining Google's scientific moonshots and safety frameworks on the threshold of artificial general intelligence.

## Atmosphere & style The presentation carries the slick, high-energy cadence of a flagship Silicon Valley keynote, balancing dizzying infrastructure metrics with playful, live-action demonstrations like compiling an operating system to run Doom and ordering coffee hands-free via audio glasses. Key executives like Sundar Pichai and Demis Hassabis project calm institutional confidence, transitioning smoothly from pragmatic enterprise productivity into expansive, philosophical forecasts about standing in the 'foothills of the singularity.' The tone oscillates between technical rigor and mass consumer excitement, punctuated by glossy video interludes, partner spotlights, and enthusiastic auditorium applause.

## Bulletin - [01:54](https://www.youtube.com/watch?v=wYSncx9zLIU&t=114s) Google marked 10 years as an AI-first company with monthly token processing reaching 3.2 quadrillion tokens, supported by annual capital expenditures climbing toward $180B–$190B and 8th-generation dual-chip TPUs. - [27:14](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1634s) Gemini 3.5 Flash and Google Antigravity 2.0 shifted software engineering toward autonomous multi-agent execution, highlighted by the generation of a functional operating system in 12 hours for under $1,000 in API credits. - [33:35](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2015s) Sundar Pichai unveiled Gemini Spark, a 24/7 cloud-hosted personal agent running on dedicated virtual machines that completes complex, asynchronous workflows across Google Workspace and external services even when devices are closed. - [58:45](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3525s) Google introduced open commerce standards including the Universal Commerce Protocol (co-steered by Amazon, Meta, Microsoft, and Stripe) and the Agent Payments Protocol to enable multi-merchant checkout and autonomous agent transactions. - [01:33:00](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5580s) Android XR expanded into physical eyewear through lightweight, screenless audio glasses designed with Samsung, Warby Parker, and Gentle Monster, scheduled for commercial release in the fall.

## Key takeaways - [01:54](https://www.youtube.com/watch?v=wYSncx9zLIU&t=114s) Google's infrastructure scale has reached 3.2 quadrillion tokens processed per month, with API volume processing roughly 19 billion tokens per minute across 8.5 million monthly active developers. - [11:17](https://www.youtube.com/watch?v=wYSncx9zLIU&t=677s) Google's projected annual capital expenditure of $180B–$190B is anchored by 8th-generation TPUs with a dual-chip strategy: TPU v8 Training for massive pre-training and TPU v8 Inference for low-latency serving. - [15:44](https://www.youtube.com/watch?v=wYSncx9zLIU&t=944s) Demis Hassabis introduced Gemini Omni as a breakthrough world model capable of simulating physical dynamics like gravity, kinetic energy, and intuitive physics, framing world modeling as an architectural prerequisite for AGI. - [23:05](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1385s) Gemini 3.5 Flash delivers frontier-class reasoning at four times the output speed of competing models and at less than half the operating price, enabling major enterprise cloud cost savings. - [25:26](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1526s) Google Antigravity 2.0 transitions developer workflows from code completion to autonomous agent orchestration, employing subagents, event hooks, and asynchronous task management. - [33:35](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2015s) Gemini Spark operates as an always-on autonomous agent hosted on Google Cloud virtual machines, executing multi-step tasks across Workspace apps and external services using the Model Context Protocol. - [45:59](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2759s) Google Search unified AI Overviews and conversational AI Mode into a single multimodal interface backed by a real-time knowledge base refreshing over 1 billion facts per minute. - [51:10](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3070s) Search now incorporates dynamic agentic coding via Antigravity to plan, compile, and render bespoke generative user interfaces and stateful mini-apps inside secure browser sandboxes. - [58:45](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3525s) The Universal Commerce Protocol (UCP) and Agent Payments Protocol (AP2) create an open, cross-industry foundation for agentic discovery, tamper-proof payment mandates, and autonomous checkout. - [01:03:20](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3800s) The Universal Cart centralizes cross-merchant shopping baskets across Search, Gemini, YouTube, and Gmail with real-time technical compatibility checks and background price drop monitoring. - [01:33:00](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5580s) Android XR intelligent audio glasses deliver zero-touch, hands-free multimodal assistance and background task execution, natively pairing with both Android and iOS devices. - [01:44:37](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6277s) DeepMind previewed Code Mender for autonomous codebase vulnerability remediation alongside scientific foundation models including Alpha Earth Foundations and Weather Next. - [01:48:56](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6536s) Demis Hassabis asserted that humanity has entered the 'foothills of the singularity,' pointing to computational virtual cell modeling at Isomorphic Labs aimed at fundamentally compressing drug discovery timelines.

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1s) Opening: Building What Matters

The keynote begins with an inspiring opening montage highlighting the rapid evolution of human creativity and technical problem-solving. Voices reflect on how far technology has progressed, driven by creators who were bold enough to believe they could solve the world's most daunting challenges—from tackling complex scientific questions and pursuing the dream of eradicating disease, to helping small business owners manage menus, marketing, and inventory. Educators who do not know how to write code describe how generative tools have unlocked completely new pathways for teaching and creating. In one personal moment, an AI assistant named Matthias explains that his brother Tommaso built him to help communicate, underscoring the film’s central message: “Let's make something that matters.”

---

[01:54](https://www.youtube.com/watch?v=wYSncx9zLIU&t=114s) Sundar Pichai on a Decade of AI-First Innovation and Token Scale

Google CEO Sundar Pichai takes the stage at the Shoreline Amphitheatre to welcome attendees and global viewers to Google I/O. Looking back on an intense year of relentless shipping and rapid technological advancement, Pichai jokes about an exaggerated reel of his personal schedule, admitting that while he hasn't reached outer space yet, a depiction of him plugging in a TPU felt quite accurate. Transitioning to a serious reflection, Pichai marks a major milestone: “It's been 10 years since we pivoted the company to be AI first.” He explains that Google’s continued momentum relies on a differentiated, full-stack approach that spans custom silicon and secure foundational infrastructure, cutting-edge research models, and consumer platforms reaching billions of people across the globe.

```mermaid flowchart TD accTitle: Google Full-Stack AI Innovation accDescr: A diagram showing the layers of Google's AI stack from silicon to user products.

subgraph Infra["Foundation Layer"] A["Custom Silicon (TPUs)"] B["Secure Cloud Foundation"] end

subgraph Core["Model Layer"] C["World-Class Research"] D["Gemini & Media Models (Lyria, Veo)"] end

subgraph Apps["Application Layer"] E["Consumer & Enterprise Products"] F["Developer Platforms & APIs"] end

Infra --> Core Core --> Apps ```

Pichai emphasizes that user stories—from students studying for exams to creators using models like Lyria and Veo—are the truest measure of progress, but points to token processing as an astonishing proxy for industry-wide adoption. Reflecting on the dizzying growth of model usage, he observes: “I never imagined I'd say quadrillion in an I/O keynote.”

- Two years ago, Google was processing 9.7 trillion tokens per month across its surfaces. - By the previous year's I/O, monthly volume reached approximately 480 trillion tokens. - Today, monthly volume has jumped sevenfold to 3.2 quadrillion tokens per month. - More than 8.5 million developers build monthly applications and experiences using Google models. - Model APIs process roughly 19 billion tokens per minute. - Over the past 12 months, more than 375 enterprise customers each processed in excess of 1 trillion tokens.

---

[05:16](https://www.youtube.com/watch?v=wYSncx9zLIU&t=316s) Scaling Gemini Across Search, the Gemini App, and Maps

Pichai details the company's product scale, noting that Google now maintains 13 distinct products with over 1 billion users each, five of which boast more than 3 billion users. A primary catalyst behind this engagement is the integration of Gemini models, beginning with Google Search. AI Overviews has expanded to over 2.5 billion monthly users. Its companion conversational experience, AI Mode—which Pichai calls Google’s “biggest upgrade to search ever”—has already surpassed 1 billion monthly users in its first year. Instead of isolated queries, Search is transforming into an ongoing conversation that delivers synthesized insights and contextual web links.

The standalone Gemini app is experiencing parallel momentum, with monthly active users climbing from 400 million at the prior I/O to more than 900 million today, while daily query requests grew more than sevenfold. Features like Personal Intelligence provide tailored assistance, and visual generation has exploded, with over 50 billion images generated to date using the team's Nano banana visual models. Conversational upgrades have also rolled out to Google Maps in its largest overhaul in a decade, allowing users to pose lengthy, multi-variable questions—such as an urgent query from a frantic parent whose child fell into a duck pond thirty minutes before a wedding, seeking a nearby walking destination to purchase a replacement dress.

---

[07:36](https://www.youtube.com/watch?v=wYSncx9zLIU&t=456s) Conversational AI in Action: Ask YouTube and Docs Live

Pichai introduces two new conversational product experiences designed to streamline how users discover information and create content:

1. **Ask YouTube**: Reimagining search across video libraries, Ask YouTube helps viewers navigate complex topics—such as transitioning a three-year-old child from a balance bike to a pedal bike. Instead of returning a wall of video thumbnails, the system delivers concise overviews, relevant instructional tips, jumps directly to the exact timestamps within recommended videos, retains conversational context for follow-up questions (like comparing handbrakes versus pedal brakes), and formats feature comparisons into easy-to-read tables. Testing is underway, with a broad rollout scheduled across the United States this summer.

2. **Docs Live**: Leveraging breakthrough technical leaps in audio models, Docs Live allows users to bypass typing precise text prompts in favor of spontaneous voice dictation. Pichai introduces a live, real-time product demonstration showing a software engineer drafting an alumni career day talk:

> "I just remembered I'm doing an alumni talk for my High School's career day tomorrow. I need to come up with some talking points to explain what I do for a living as a software engineer, but I'm not really sure where to start. Oh, actually, can you just pull my resume from drive. Although that might be boring. Maybe can you come up with some funny analogies so it'll be more of an engaging talk for the students. Oh, and also, I think the school sent me an email... I think the subject is something like career day logistics. Maybe just grab the details from there, throw them at the top of the doc so I know where to go and what time to get there."

Within seconds, Docs Live accesses the user's Google Drive and Gmail, extracts the relevant details, generates an engaging draft, and then responds to a follow-up voice command to reformat the analogies into a clean comparison table and bold an introductory note about the speaker's brother. Docs Live will roll out to Gemini Pro and Ultra subscribers this summer, with identical voice editing capabilities coming to Gmail and Google Keep.

---

[11:17](https://www.youtube.com/watch?v=wYSncx9zLIU&t=677s) Massive Infrastructure Investments and 8th-Gen Dual-Chip TPUs

Delivering these AI experiences at global scale while simultaneously supporting millions of external developers and enterprise workloads demands unprecedented investments in data center infrastructure and silicon. Pichai reveals the dramatic escalation in capital expenditure over the past four years: in 2022, Google's annual CapEx stood at $31 billion; for the current year, capital expenditures are projected to increase approximately sixfold, reaching between $180 billion and $190 billion.

Marking ten years since Google introduced its very first commercial Tensor Processing Unit (TPU) on the I/O stage, Pichai highlights the 8th generation of TPUs recently unveiled at Cloud Next. For the first time, Google has adopted a dual-chip strategy featuring specialized microarchitectures tailored separately for training and inference workloads:

- **Training Chip (TPU v8 Training)**: Engineered specifically for massive-scale pre-training workloads, delivering nearly three times the raw computing performance of the previous generation. - **Inference Chip (TPU v8 Inference)**: Optimized exclusively for low-latency, high-throughput model execution and serving. - **Global Training Clusters via JAX and Pathways**: Google’s training topology is no longer restricted by the physical bounds of a single mega-data center; rather, training runs are seamlessly orchestrated across distributed sites worldwide, enabling model builders to train larger and more capable frontier models in weeks instead of months.

[13:03](https://www.youtube.com/watch?v=wYSncx9zLIU&t=783s) TPU Latency, Flash Speed, and Sustainable Scaling

Google emphasizes that speed has been engineered into every layer of its systems, drawing on decades of infrastructure experience. As the presenter explains, “We have dramatically improved speed at every step because if we learn anything in 27 years of working on search, it's that latency matters.” To illustrate the real-world impact of these hardware and model optimizations, a live demonstration showcases an upcoming Flash model tasked with generating a playable Chrome Dino game. The response stream appears on screen almost instantaneously, reaching an output throughput of nearly 1,500 tokens per second—making the code generation faster than the time it took to type out the prompt.

Alongside sheer generation speed, Google frames sustainable scaling as a core engineering priority. The underlying processing chips feature improved energy efficiency, yielding up to two times better performance per watt. To provide lighthearted comic relief about the scale of compute running behind the scenes, an animated sketch follows conversational TPU processors preparing for the conference. In the breakroom banter, one chip mentions folding proteins across rare oncology datasets while another simulated 50 years of climate models, contrasting with a third chip that generated an image of a pug dressed as an accountant. The sequence concludes as a coach chip rallies the units—urging "Timmy TPU" to clear cache and dry fans before tackling trillions of jobs ahead of I/O.

- Live demo generation rate reached nearly 1,500 tokens per second on an upcoming Flash model. - Hardware updates deliver up to 2x better performance per watt for sustainable data center scaling.

---

[15:44](https://www.youtube.com/watch?v=wYSncx9zLIU&t=944s) Demis Hassabis Introduces Gemini Omni and World Models

Returning to the main stage, the presentation pivots to three foundational areas of technical progress: models, coding, and agents. The discussion begins with the evolution of world models, representing an architectural leap from basic statistical text prediction to full-scale environmental simulation. Demis Hassabis, CEO of Google DeepMind, takes the stage to frame world modeling as an essential milestone toward artificial general intelligence. Reflecting on the acceleration of the field over the previous twelve months, Demis notes: “We now have agents that can plan and act on our behalf. And artificial general intelligence is just a few years away.”

```mermaid flowchart TD accTitle: Architecture of Gemini Omni World Model accDescr: Diagram showing how multimodal foundation models combine with generative media to power Gemini Omni for simulation, editing, and video generation.

subgraph Inputs["Inputs & Foundation"] A["Any Input Modality\n(Text, Video, Image, Audio)"] B["Gemini World Knowledge\n& Reasoning"] C["Generative Media Stack\n(Veo, Imagen, Genie)"] end

subgraph Core["Gemini Omni Engine"] D["Gemini Omni\n(Intuitive Physics & Simulation)"] end

subgraph Outputs["Capabilities & Outputs"] E["Physical Reality Simulation\n(Gravity, Kinetic Energy)"] F["Conversational Video Editing\n(Iterative Scene Morphing)"] G["Multi-Modal Artifacts\n(Claymation, Explainer Media)"] end

A --> D B --> D C --> D D --> E D --> F D --> G ```

Demis officially unveils Gemini Omni, a model built to ingest any input and generate any output. It combines the reasoning and broad world knowledge of the Gemini architecture with Google's generative media technologies (including Veo and Genie) to understand intuitive physics, kinetic energy, and gravity. To illustrate its pedagogical and creative capabilities, Demis demonstrates a prompt asking for a claymation explainer of protein folding, showing amino acid chains assembling into alpha helices and beta sheets.

Crucially, Omni moves beyond one-shot generation into iterative, conversational video modification. Users can upload personal mobile footage—such as a selfie or an evening walk—and adjust styles, insert elements, or transform objects (like expanding a hand-drawn circle into a cosmic black hole) through natural dialogue. Demis confirms the release of Gemini Omni Flash across Google products, with Gemini Omni Pro planned for near-term release.

- **Kinetic simulation**: Integrates physics principles to govern motion, weight, and environmental interactions. - **Conversational video manipulation**: Enables multi-turn, prompt-based edits on user-provided source video. - **Gemini Omni Flash**: Available immediately across products, serving as the lightweight entry point to the Omni family.

---

[20:47](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1247s) SynthID Expansion, Content Credentials, and Deepfake Detection

Sundar Pichai steps back onto the stage to address the safety and authenticity challenges stemming from increasingly realistic generative media. He highlights research showing that humans can correctly identify high-grade deepfake videos only about 25 percent of the time. To counteract deceptive media, Google is scaling SynthID, its invisible digital watermarking technology introduced three years prior. Over its lifespan, SynthID has embedded watermarks across more than 100 billion images and videos, as well as 60,000 years' worth of audio content.

```mermaid flowchart LR accTitle: Content Provenance and Verification Pipeline accDescr: Pipeline illustrating how captured and edited assets receive SynthID watermarks and Content Credentials for verification in Search and Chrome.

subgraph CaptureAndEdit["Media Creation & Editing"] A["Pixel Camera Capture"] --> B["Google Photos / GenAI Editing"] B --> C["SynthID Watermark &\nC2PA Content Credentials"] end

subgraph Verification["Consumer Surface Inspection"] C --> D["Circle to Search"] C --> E["Chrome Right-Click"] D --> F["Provenance Card:\nOrigin & Tool History"] E --> F end ```

Google is now coupling SynthID with Content Credentials verification directly inside consumer tools like Google Search and Chrome. Users can invoke Circle to Search or right-click any web asset to query its synthetic origins, receiving an immediate verification card displaying capture source (e.g., Pixel camera) and editing history (e.g., Google Photos generative tools). Pichai humorously shows a viral social media image of himself eating a hamburger, dryly noting, “It's obviously fake. I don't eat hamburgers.” Because provenance systems require broad ecosystem buy-in, Pichai announces significant cross-industry partnerships: following NVIDIA's adoption the prior year, OpenAI, Kakao, and ElevenLabs have also officially agreed to adopt SynthID.

- **Scale of SynthID**: Over 100 billion visual assets (images and videos) and 60,000 years of audio marked to date. - **Detection latency**: Integrates directly into Circle to Search and Chrome desktop context menus. - **Industry partners**: OpenAI, Kakao, and ElevenLabs join NVIDIA in deploying SynthID watermarking standards.

---

[23:05](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1385s) Gemini 3.5 Flash and Internal Agent Workflows

Pichai transitions to developer infrastructure and model upgrades within the Gemini 3 family. Reflecting on the rapid uptake of the architecture, he notes that developers have treated Flash models as their everyday development engine while utilizing Pro variants for deep reasoning. To push these capabilities further, Google introduces Gemini 3.5 Flash, an iteration specifically tuned for agentic coding, multi-step long-horizon reasoning, and complex execution paths. Pichai emphasizes its balanced profile: “When looking at output tokens per second, it's four times faster than other frontier models, and it's incredible delight to use.”

Benchmark figures presented on stage reveal that Gemini 3.5 Flash outperforms the earlier 3.1 Pro model across nearly all evaluations, displaying notable leaps on coding and GDP-Val (a benchmark tracking economically valuable, real-world workplace tasks). Internally, Google has deployed 3.5 Flash inside its dedicated agent-first development platform, Antigravity. Internal developer token consumption climbed from 500 billion tokens per day in March to over 3 trillion tokens per day by May, creating an aggressive data feedback loop used to continuously refine agentic responsiveness and code generation accuracy.

- Outperforms Gemini 3.1 Pro across standard benchmarks while operating four times faster than competing frontier models. - Demonstrates large performance gains on the GDP-Val benchmark for economically impactful multi-step tasks. - Internal Google developer usage grew from 0.5 trillion tokens daily in March to over 3 trillion tokens daily ahead of I/O.

---

[25:26](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1526s) Varun Mohan Unveils Google Antigravity 2.0

Varun Mohan takes the stage to walk through the transformation of software engineering from simple code-completion tools to autonomous acting agents. He frames this shift as a fundamental democratization of development: “We've moved beyond AI tools that help us write to agents that help us act. These agents have lowered the barrier to development so much that anyone can be a builder, even busy CEOs.” In an aside, Mohan shares that Sundar Pichai used Antigravity the preceding week to locate and patch a bug directly inside Google's monolithic codebase.

Following the initial launch of the Antigravity IDE preview in November, Google is turning developer feedback into a full product expansion. Antigravity now features a dedicated Command Line Interface (CLI), an official SDK, native speech integration driven by Gemini audio models, and pipeline hooks for Android Studio, Firebase, and Google AI Studio.

At the center of the announcement is Antigravity 2.0, reconstructed as a standalone, agent-optimized desktop application. Instead of treating agent interactions as a side panel in a traditional code editor, Antigravity 2.0 is built from the ground up around agent conversations, persistent agent-generated artifacts, and multi-agent orchestration engines.

- **Antigravity CLI & SDK**: Command-line tools and software development kits for running agentic tasks outside the traditional editor GUI. - **Native audio**: Integrates Gemini voice models directly for hands-free, conversational debugging and task delegation. - **Standalone architecture**: Antigravity 2.0 transitions from an IDE extension/variant into an autonomous desktop workspace focused on multi-agent execution.

[27:14](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1634s) Autonomous OS Development with Google Antigravity and Gemini 3.5 Flash

Google's internal teams have been operating on an unapologetically "agent-first" paradigm to unlock major engineering value. At the center of this transformation is the Antigravity agent harness, an invisible framework designed to allow Gemini models to execute real-world computing workflows. The framework has recently gained critical core primitives, including subagents, event hooks, and asynchronous task management. All of these capabilities are anchored by Gemini models, with Gemini 3.5 Flash specifically optimized to pair directly with the Antigravity harness.

Curious to test the absolute boundaries of what autonomous multi-agent systems could achieve, Google engineers assigned the system a notoriously brutal software engineering challenge: creating a fully functioning operating system from an empty directory. Over a 12-hour period, Antigravity independently decomposed the objective into an overarching technical plan, distributed the work across parallel subagents, and iteratively authored, executed, and refined test suites. The autonomous team wrote every single line of code, implementing core operating system components such as the process scheduler, memory management subsystem, and virtual file system.

```mermaid flowchart TD accTitle: Antigravity Autonomous OS Generation Architecture accDescr: Flowchart illustrating how Antigravity coordinates parallel subagents to generate a functional operating system from an empty project.

subgraph Planning ["1. Task Decomposition"] A["Empty Project Prompt: Build an OS"] --> B["Antigravity Agent Harness"] B --> C["Cohesive Architectural Plan"] end

subgraph Execution ["2. Parallel Execution & Testing"] C --> D["93 Autonomous Subagents"] D --> E["15,000+ Model Requests<br/>2.6B Tokens Processed"] E --> F["Automated Test Generation & Iteration"] end

subgraph Result ["3. Completed Operating System"] F --> G["Kernel & Process Scheduler"] F --> H["Memory Management"] F --> I["File System"] end ```

The scale and cost efficiency of this achievement highlighted a major leap forward from prior model generations. As Varun explained: “This was not possible on Gemini 3.1 Pro, but thanks to the performance and cost efficiency of Gemini 3.5 Flash, building an entirely functional operating system consumed less than $1,000 of API credits.” Rather than merely building an isolated software application, the agents produced an entire operational foundation capable of loading drivers, handling hardware interrupts, and hosting standard user-space applications.

- **Subagents deployed:** 93 parallel subagents working concurrently over a 12-hour span. - **Compute scale:** Over 15,000 model requests processing 2.6 billion tokens. - **Development cost:** Under $1,000 in total API credit consumption. - **Architectural depth:** Entire codebase generated from scratch, covering process scheduling, memory virtualization, and file systems.

---

[29:34](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1774s) Live Demo: Antigravity 2.0, Driver Generation, and Playing Doom

To prove the operating system was genuinely operational rather than a simulated mock-up, Varun switched directly into an active terminal window running on the Antigravity-generated OS. Demonstrating a functional OS on stage can be dry, so he began with the classic terminal command `sl`—a playful utility triggered when engineers mistype `ls`. Right on cue, an ASCII steam locomotive chugged across the terminal window sporting a custom Antigravity logo. However, Varun noted that in computing lore, no operating system is truly complete until it can run the classic 1993 game *Doom*. Attempting to launch the game failed immediately because the custom OS lacked the required video and keyboard drivers.

To resolve the issue in real time, Varun switched into the newly redesigned Antigravity 2.0 interface. The user experience is structured strictly around agent interactions, featuring active projects and conversation threads pinned to the left-hand navigation panel. While the agent worked on the driver issue in the background, Varun toured an existing research thread where an agent had analyzed *Doom* trivia, generating live analytical plots on the right-hand panel, code-derived charts, structured data tables, and an infographic produced with Nano Banana Pro. Turning back to the driver prompt, Antigravity had researched the hardware specifications, authored over 100 lines of low-level driver code, recompiled the operating system kernel, and installed the updates. Returning to the live terminal, Varun launched the game again, and *Doom* loaded smoothly to widespread audience cheers.

Varun emphasized that this breakthrough extends far beyond retro gaming demonstrations. Google's engineering teams have subjected Antigravity to complex real-world software requests, including writing an entire photo editing suite, building a real-time messaging application, and delivering a multi-user collaborative workspace. Across all these tests, Varun observed: “Multi-day engineering efforts are collapsing into hours, if not minutes.” To support this token-heavy agentic workflow, Google tuned Gemini 3.5 Flash inside the platform to run 12 times faster than typical frontier models. Antigravity 2.0 launched globally for all developers immediately following the keynote presentation.

---

[33:35](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2015s) Sundar Pichai on Gemini 3.5 Flash Economics and Announcing Gemini Spark

Returning to the stage, Sundar Pichai marveled at the fact that Varun's entire operating system was assembled by a team of subagents in just 12 hours for less than $1,000. He underlined that the core promise of Gemini 3.5 Flash is delivering frontier-class reasoning and execution capabilities at less than half the operating price of competing frontier models. Sundar observed that enterprise token consumption has escalated dramatically, noting that many organizations had already exhausted their entire annual AI budgets by May.

To quantify the financial impact, Sundar pointed to Google Cloud's largest corporate customers, who collectively process roughly 1 trillion tokens per day. As Sundar calculated: “If they shifted 80% of their workloads from other frontier models to 3.5 Flash, they'd save over $1 billion annually.” Gemini 3.5 Flash was made globally available across Google's developer APIs and consumer products immediately, with the flagship Gemini 3.5 Pro model slated for release the following month.

```mermaid flowchart LR accTitle: Gemini Spark Cloud Architecture accDescr: Diagram illustrating how Gemini Spark operates continuously on Google Cloud virtual machines to automate tasks across first-party and third-party tools.

subgraph Client ["Client Access"] User["User Direction"] --> UI["Gemini App / Chat / Email"] end

subgraph Cloud ["Google Cloud Infrastructure"] UI --> VM["Dedicated Virtual Machines (24/7)"] VM --> Engine["Gemini 3.5 & Antigravity Harness"] end

subgraph Integrations ["Ecosystem & Tools"] Engine --> GoogleApps["Google Workspace (Docs, Sheets, Slides, Gmail)"] Engine --> MCP["Third-Party Tools (via MCP)"] end ```

Beyond enterprise and developer workflows, Sundar shifted focus toward bringing agentic computing safely and securely to everyday consumers. He unveiled Gemini Spark, a personal AI agent integrated directly inside the Gemini mobile and desktop applications. Spark is engineered to handle complex, open-ended digital chores completely autonomously in the background. As Sundar highlighted: “It runs on dedicated virtual machines on Google Cloud and it's 24/7, and yes, you can close your laptop.” Spark is orchestrated by Gemini 3.5 and the Antigravity agent harness, enabling it to execute multi-step workflows across Google Workspace and external services via the Model Context Protocol (MCP).

- **Enterprise scale:** Leading Google Cloud enterprise accounts process roughly 1 trillion tokens daily. - **Cost reduction:** An 80% workload migration to 3.5 Flash unlocks more than $1 billion in annual savings. - **Product availability:** Gemini 3.5 Flash launched globally on keynote day; Gemini 3.5 Pro ships the following month. - **Gemini Spark infrastructure:** Runs continuously on dedicated Google Cloud virtual machines without requiring the user's local device to stay powered on. - **Protocol support:** Connects across Google first-party tools and third-party integrations via the Model Context Protocol (MCP).

---

[36:26](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2186s) Josh Woodward Demonstrates Gemini Spark in Action

Josh Woodward took the stage to demonstrate how Gemini Spark operates inside a freshly redesigned Gemini interface. Opening the application, Woodward displayed Spark’s central mission-control dashboard, which surfaces active and pending asynchronous jobs so users can monitor progress at a glance. To show its day-to-day utility, he submitted a prompt asking Spark to compile recent product launches and team accomplishments from the past week into an internal update email, specifying the use of a custom skill called "Ghostwriter." Woodward explained that Ghostwriter is a personalized behavioral style guide he configured so that machine-generated communications match his exact voice, tone, and formatting preferences. Spark immediately began querying Google Docs, Gmail threads, and chat logs across the defined seven-day window to synthesize the update.

Woodward then pivoted to a complex personal scenario: organizing an upcoming neighborhood block party. His prompt instructed Spark to collect attendee RSVPs, track what food and supplies each family volunteered to bring, and draft polite reminders to neighbors who had not yet responded. Spark parsed through chaotic email replies and structured the information into a live RSVP tracker inside Google Sheets. Because Spark maintained an active integration with Gmail, the spreadsheet automatically updated in real time whenever an email arrived, such as an RSVP confirmation from a neighbor named L. Thompson.

The agent took initiative beyond basic administrative organization. It prepared ready-to-send draft reminder messages inside Gmail so Woodward could review and dispatch them with one click. It also assembled a full "hype deck" presentation in Google Slides, pulling in visual details such as a massive bounce house planned for the street's cul-de-sac. Demonstrating proactive contextual awareness, Spark flagged an important administrative constraint it discovered during its background research: the local homeowners association rules prohibited setting up party equipment on the street prior to Friday afternoon, June 5. All of these background tasks ran independently on Google Cloud, keeping the user firmly in the loop while handling the tedious logistics end-to-end.

[39:16](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2356s) Mobile Multitasking and Asynchronous Brain Dumping in Gemini Spark

After demonstrating Gemini Spark on a laptop, Josh shows how the agent seamlessly transitions to mobile devices across both Android and iPhone, syncing all ongoing tasks in real time. Because Spark is designed to let busy users offload mental clutter, Josh demonstrates voice-dictating an unfiltered stream of three disparate personal and professional requests in a single breath: finding upcoming meetings with Sundar and turning them hot pink in Calendar so they aren't missed; drafting a warm block party invitation note to a new neighbor named John and his family; and generating an end-of-school-year family to-do checklist for his kids, sorted by deadline and priority.

```mermaid flowchart TD accTitle: Gemini Spark Voice-to-Action Flow accDescr: Diagram showing how an unstructured voice brain dump is captured, parsed, split into tasks, and executed in the background across devices.

subgraph Input ["User Input"] A["Spoken Voice Dictation<br/>(Multiple tasks at speech speed)"] end

subgraph Spark ["Gemini Spark Processing"] B["Single Unified Thread"] C["Context Parser & Task Deconstructor"] B --> C end

subgraph Execution ["Asynchronous Background Execution"] D1["Calendar Agent:<br/>Color-code Sundar meetings hot pink"] D2["Gmail/Docs Agent:<br/>Draft block party invite to neighbor John"] D3["Drive/Docs Agent:<br/>Prioritized end-of-school checklist"] end

A --> B C --> D1 C --> D2 C --> D3 ```

Rather than requiring step-by-step form filling, Spark captures the entire context as fast as he speaks, initializes everything under one parent thread, and deconstructs it into separate background jobs. Josh then puts his phone down on the stage table, highlighting that the system works autonomously without requiring the user to remain tethered to the screen.

- Cross-platform syncing ensures tasks started on desktop immediately mirror onto Android and iOS devices. - Single-thread multi-task parsing handles compound voice requests without forcing conversational pauses between instructions. - Autonomous background execution allows users to lock or put away their device while agents carry out multi-step actions.

---

[41:02](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2462s) Gemini Spark Rollout, Pricing Tiers, and the Agentic Era

Josh outlines the rollout timeline and revised pricing tiers designed to put agentic assistance into broader distribution while maintaining safety boundaries. Gemini Spark begins rolling out immediately to trusted testers, followed by a beta release for Google AI Ultra subscribers the following week. To make access more affordable, Google is introducing a new $100-per-month Ultra tier and cutting the price of its top-tier, maximum-limits Ultra tier from $250 to $200 per month. Looking ahead, Josh teases two upcoming milestones: Spark operating natively inside Google Chrome as an agentic browser capable of navigating and executing tasks across the open web, and Android Halo, a dedicated mobile home base for managing device agents arriving later in the year.

Returning to the stage, Sundar Pichai reflects on the significance of this transition:

> “We are firmly in our agentic Gemini era.”

Pichai emphasizes that Gemini Spark represents the first tangible manifestation of pairing Gemini 3.5 models with Google Antigravity, transforming Google’s products from passive answer engines into active problem-solving collaborators. He notes that while people have experimented with early agents, building experiences that are simultaneously effortless, secure, and helpful remains a hard challenge, paving the way for search—Google's original moonshot—to undergo its biggest evolution yet.

- **Rollout Schedule:** Trusted testers access Gemini Spark immediately; Google AI Ultra subscribers receive beta access the following week. - **Ultra Tier Pricing:** New Ultra plan launched at $100/month; top-tier plan reduced from $250/month to $200/month. - **Chrome Integration:** Native agentic web browsing launching later in the summer. - **Android Halo:** Dedicated phone-based command center for AI agents launching later in the year.

---

[43:42](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2622s) The Evolution of Search: Upgrading AI Mode to Gemini 3.5

Liz Reid takes the stage to discuss the ongoing transformation of Google Search, charting the progression from foundational ranking to AI Overviews and the debut of AI Mode. AI Mode, upgraded to run on Gemini 3.5, has crossed one billion monthly active users, with query volume doubling every quarter since its introduction. Reid explains that this expanded capability has fundamentally changed query behavior: overall search volume reached an all-time high in the previous quarter because users now treat the search box as a conversational partner, submitting highly nuanced, multi-constraint prompts rather than fragmented keywords.

Instead of searching for broad phrases like "nearby hikes," users routinely ask complex, composite questions such as building a full-day hiking trip itinerary nearby with scenic views, dog-friendly trails, convenient parking, and a lunch stop. Reid describes the new paradigm where AI is no longer a bolt-on feature, stating:

> “Google Search is AI search through and through.”

This architecture operates on an index refreshing over one billion facts per minute, ingesting billions of newly published web pages daily, and linking them to human perspectives across the open web.

- **User Scale:** AI Mode has surpassed 1 billion monthly active users. - **Query Trajectory:** AI Mode queries have more than doubled quarter-over-quarter, propelling overall Google Search queries to an all-time high. - **Model Foundation:** AI Mode is upgraded to Gemini 3.5. - **Data Freshness:** Google's real-time knowledge base updates over 1 billion facts every minute.

---

[45:59](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2759s) The Reimagined Search Box and Unified AI Search Experience

Reid announces the most substantial redesign of Google's iconic search box since its debut over 25 years ago. The new search box is no longer a static, one-line text field; it expands dynamically as curiosity unfolds, using AI-driven formulation assistance that extends far beyond conventional autocomplete by surfacing latent nuances and contextual angles users might not have considered. The input interface natively supports multimodal inputs—including text, images, uploaded files, and video clips—and reasons across all of them simultaneously.

In parallel, Google is unifying AI Overviews and AI Mode into a single continuous interface across desktop and mobile globally. Rather than forcing users into separate tabs or tools, queries transition fluidly from an initial result summary to conversational follow-ups while preserving complete conversational state and memory. As exchanges deepen, the citations, hyperlinks, and web sources update dynamically to surface increasingly granular source material, ensuring users retain direct access to the wider web.

- **Input Modalities:** Text, images, documents/files, and video are parsed concurrently within the primary search box. - **Global Availability:** The unified AI Overviews and AI Mode experience rolls out immediately worldwide across desktop and mobile devices. - **Context Preservation:** Follow-up questions maintain thread history and context without resetting the search session.

---

[47:41](https://www.youtube.com/watch?v=wYSncx9zLIU&t=2861s) 24/7 Information Agents and Dynamic Agentic Coding in Search

Expanding on Google's agentic roadmap, Reid reveals that Google Search is introducing persistent "information agents" that operate continuously in the background on the user's behalf. Working alongside Gemini Spark, these autonomous workers evaluate user criteria, determine urgency, establish web triggers, and tap into specialized live data feeds. When noteworthy changes occur, the agent synthesizes an update that strips out noise, alerts the user, and directs traffic to relevant publishers, specialized research platforms, news sites, and social communities.

To illustrate their breadth, Reid details three practical examples: - **Financial Research:** Monitoring biotech companies with price-to-earnings (P/E) ratios under 15, positive cash flows, and low debt, triggering alerts using real-time market data the moment prices or fundamentals shift. - **Apartment Hunting:** Ingesting an exhaustive brain dump of living preferences—such as neighborhood boundaries, natural light, and move-in dates—and continuously scanning rental listings, forums, and social posts. - **Sneaker Drops:** Tracking athlete collaboration announcements across blogs, forums, and Google’s Shopping Graph to ensure fans never miss limited-edition releases.

Closing her segment, Reid notes that these information agents will launch during the summer, forming the vanguard of a broader suite of search agents capable of finding, vetting, booking, and purchasing items. She then introduces Robby Stein to showcase how agentic coding allows Search to program and render bespoke interfaces on the fly, tailoring the visual output format—whether product cards, comparative data tables, interactive charts, or rich media—to the specific intent of any query.

[51:10](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3070s) Bringing Antigravity and Generative UI to Google Search

Google is integrating its Antigravity platform alongside the agentic coding capabilities of Gemini 3.5 Flash directly into Google Search. Instead of relying on static result pages, Search can now plan, design, and compile custom user interfaces on the fly to match the exact requirements of a query. The system constructs dynamic layouts, builds custom interactive widgets, fans out research across the web, and executes code within a sandboxed, containerized environment powered by Antigravity—the identical runtime environment demonstrated earlier by Varun to construct an entire operating system.

```mermaid flowchart TD accTitle: Generative UI Architecture in Search accDescr: Flow showing how Gemini 3.5 Flash and Antigravity turn a user query into an interactive custom UI.

subgraph UserStage["User Interaction"] Q["User Query"] end

subgraph ReasoningStage["Gemini 3.5 Flash Engine"] Plan["Response Planning & UI Design"] Research["Fan-out Web Research"] end

subgraph ExecutionStage["Antigravity Coding Harness"] Container["Secure Containerized Sandbox"] CodeGen["Code Generation & File Execution"] end

subgraph OutputStage["Dynamic Interface"] Widget["Custom Interactive Component / Visual"] end

Q --> Plan Plan --> Research Plan --> Container Container --> CodeGen CodeGen --> Widget ```

To demonstrate the educational power of real-time interface generation, a college astrophysics query is posed: “how do black holes affect space time.” Rather than returning traditional links or a simple text summary, Search returns an interactive visual directly within the AI Overview. When the user follows up with a more complex request—“now show me how two orbiting objects like binary black holes create gravitational waves”—Search dynamically rewrites the interface in real time. The user is presented with adjustable parameters such as orbital separation and mass ratio, visualizing how wave patterns warp space-time as the smaller black hole spirals inward, while surfacing deeper primary literature like the LIGO discovery papers below the widget.

- **Underlying Engine:** Gemini 3.5 Flash combined with the Antigravity agentic coding harness. - **Security & Execution:** Reads, writes, and executes code inside a secure containerized environment. - **Rollout Schedule:** Generative UI powered by Antigravity rolls out to Search globally during the summer free of charge.

---

[54:07](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3247s) Stateful Mini-Apps and the Personal Weekend Planner Demo

Moving past one-off informational queries, Search is expanding into multi-session, stateful mini-apps, dashboards, trackers, and planning tools. When users encounter long-running or recurring tasks—such as coordinating a home move, organizing a wedding, or planning recurring family outings—Search proactively generates fully functional custom software experiences that persist over time.

During a live demonstration of a user searching for family weekend activities, AI Mode proactively offers to assemble a customized weekend planner. Opening a "peek under the hood" window exposes the system's real-time thinking steps and code generation streams. Because the user connected Google Calendar, Gmail, and Google Photos, Search leverages personal intelligence: it reads receipts, checks family schedules, and notes specific household context—such as having two children, one who loves animals and an older child who is learning chess. The compiled planner arranges events chronologically, factors in real-time driving times and local weather forecasts, selects a chess meetup, and locks in the Happy Hollow Park and Zoo. Because the afternoon already contains an entry to watch a sporting event with a friend, Search leaves that block open, populating dinner reservations beneath an embedded Google Map.

The interface remains entirely editable via natural language prompts. When the presenter requests to add a recurring Friday date night and relocate the map widget to the top of the view, Search re-evaluates the layout and updates the components immediately. The resulting interactive mini-app is fully shareable via a simple link: the presenter previews how the planner renders on his wife Danielle’s smartphone, jokes about pending feedback from home, and commits the finalized itinerary across all family calendars with a single tap.

- **Personal Intelligence Integrations:** Secure opt-in linking for Gmail, Google Photos, and Google Calendar. - **Stateful Persistence:** Mini-apps remember prior constraints, adjust layouts conversationally, and allow ongoing revisions week over week. - **Availability:** Custom stateful experiences roll out in the coming months, beginning with subscribers.

---

[58:45](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3525s) The Foundation for Agent Commerce and the Universal Commerce Protocol

Vidhya Srinivasan takes the stage to address the rapid evolution of commerce, noting that shoppers turn to Google over one billion times every single day. Powering these interactions is the Google Shopping Graph, which now tracks more than 60 billion constantly updated product listings. By uniting this catalog scale with multimodal Gemini models, Google is establishing a framework for autonomous, agentic shopping built on three foundational pillars: open communication, delegated payments, and cross-merchant state management.

The first building block is the Universal Commerce Protocol (UCP). Vidhya Srinivasan emphasizes its role as an industry-wide open standard:

> “UCP does for agent commerce what HTTP did for the web.”

Developed in collaboration with founding ecosystem partners, UCP provides a unified, open-source language allowing autonomous agents, retailers, and logistical systems to coordinate every phase of the purchasing lifecycle—including exploratory product research, checkout transactions, and post-purchase shipment tracking. Highlighting broad industry alignment across tech giants, Srinivasan observes:

> “We co-developed and launched UCP with an incredible group of founding partners and recently welcomed Amazon, Meta, Microsoft, Salesforce, and Stripe. They're going to help steer this as the open standard. It may very well be the first time we've all agreed on something.”

- **Shopping Graph Scale:** Over 60 billion real-time listings integrated with Gemini models. - **Steering Partners:** Amazon, Meta, Microsoft, Salesforce, and Stripe join Google in standardizing UCP. - **Vertical & Geographic Expansion:** UCP is broadening into hospitality, local food delivery platforms, and YouTube, with expanded experiences launching in Canada, Australia, and the United Kingdom in the coming months.

---

[01:01:10](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3670s) Agent Payments Protocol (AP2) and the Universal Cart

Addressing the primary consumer anxiety surrounding AI-driven shopping—the fear of an autonomous agent making unapproved, errant purchases—Google introduced the Agent Payments Protocol (AP2). As Vidhya Srinivasan explains:

> “It's designed so your AI agent can securely make payments on your behalf, but always under your control.”

AP2 operates on two strict pillars: boundary definition and transaction accountability. Users configure specific guardrails by dictating preferred retailers, specific SKU types, and strict budgetary ceilings; the agent is only permitted to execute the final transaction when all predefined criteria align. To prevent unauthorized activity, AP2 establishes a transparent, cryptographically verifiable link between the shopper, merchant, and payment processor using privacy-preserving techniques. Transactions issue tamper-proof digital mandates, producing an unalterable digital paper trail that guarantees both merchant and consumer reference the identical purchase ledger during dispute or return inquiries while payment credentials remain completely shielded.

Concluding the commercial framework, Srinivasan introduces the third core pillar: the Universal Cart. Designed as a truly intelligent, multi-merchant container, the Universal Cart unifies shopping baskets across disconnected platforms, vendors, and services into a single persistent checkout experience.

- **AP2 Security Model:** Employs tamper-proof digital mandates, privacy-preserving credential shielding, and unified digital paper trails for returns. - **AP2 Deployment:** Integrates across Google products in the coming months, debuting first within Gemini Spark. - **Universal Cart:** Consolidates multi-vendor item discovery and checkout into an interoperable basket across separate digital storefronts.

[01:03:20](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3800s) Intelligent Shopping with the Universal Cart

The presentation highlights a reimagined e-commerce experience embedded directly across Google's ecosystem. Users can seamlessly add items to their universal shopping cart while browsing Search, interacting with the Gemini app, watching YouTube videos, or reading emails in Gmail. Once an item is placed in the cart, Gemini models work autonomously in the background to track price histories, monitor sudden price drops, scout active deals, and trigger back-in-stock notifications. Beyond tracking numbers, the cart applies intelligent reasoning to technical purchases. In an illustrative scenario where a user selects components for their first custom PC build, the cart notices that the chosen processor requires a completely different motherboard socket than the selected board, proactively warning the buyer and recommending a compatible alternative before checkout.

```mermaid flowchart TD accTitle: Universal Cart Architecture and Intelligence accDescr: Diagram illustrating how Gemini powers the Universal Cart across Google surfaces to provide automated savings, compatibility checks, and one-tap checkout.

subgraph Surfaces["Google Surfaces"] S1["Google Search"] S2["Gemini App"] S3["YouTube & Gmail"] end

subgraph Engine["Gemini Background Engine"] E1["Price History & Drop Tracking"] E2["Compatibility Reasoning Socket Checks"] E3["Google Wallet Perk & Offer Optimization"] end

subgraph Checkout["Checkout Options"] C1["Google Pay Instant In-App Checkout"] C2["Direct Cart Transfer to Retailer Site"] end

Surfaces --> Engine Engine --> Checkout ```

The cart also taps into Google Wallet to surface hidden savings by matching purchases against individual card benefits and merchant rewards, demonstrated through a Target offer linked to items added earlier in the week. Checkout is streamlined through the Universal Commerce Protocol (UCP), allowing consumers to complete purchases in a few taps using Google Pay or seamlessly transfer the entire basket to the retailer's native website.

- The universal cart begins rolling out in the United States across Search and the Gemini app during the summer, expanding to YouTube and Gmail later in the year. - As the speaker summarized the broader vision: “The building blocks we showed you today won’t just create the foundation for e-commerce, they will transform the way you shop on Google, making it more powerful, more intelligent. And a lot more fun.”

---

[01:05:51](https://www.youtube.com/watch?v=wYSncx9zLIU&t=3951s) Closing the Search Keynote: The Era of Agentic Search

Liz Reid takes the stage to bring the Search keynote to a close, framing the shift as a fusion of traditional search infrastructure with proactive, multimodal artificial intelligence. Google Search is positioned no longer as a static index of links, but as an active layer capable of generating bespoke interfaces, executing multi-step agentic tasks around the clock, and orchestrating complex commercial transactions. As she concluded her remarks: “So in this chapter of AI Search, Google can help you find, understand, build, do anything. So go ahead, just ask.” Her send-off transitions into a musical interlude featuring Depeche Mode's classic anthem “Just Can’t Get Enough” before shifting to consumer AI applications.

---

[01:08:01](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4081s) Gemini's Growth and the Neural Expressive Redesign

Josh Woodward opens the Gemini segment by detailing the rapid scale and creative adoption of the Gemini application over the prior year. More than 900 million active monthly users now use Gemini, supported by the integration of personal intelligence that links user context across Gmail, Google Photos, and other workspace apps. Woodward jokes about using personal intelligence as a personal thought partner, navigating choices like “finding the right set of tires for our super cool minivan that's going to be at that block party.” Beyond everyday tasks, NotebookLM integration has expanded dramatically, generating over 1.5 billion notebooks, audio podcasts, and slide presentations. Supported by the newly deployed Gemini 3.5 Flash model, the Gemini app now operates across more than 230 countries and over 70 languages.

```mermaid flowchart LR accTitle: Neural Expressive Interface Generation accDescr: Flow diagram showing how Gemini processes prompts and outputs dynamic visual interfaces instead of static text blocks.

U["User Prompt / Query"] --> M["Gemini 3.5 Flash Model"] subgraph UI["Neural Expressive Output"] M --> D1["Interactive Generative Images"] M --> D2["Drill-Down Information Cards"] M --> D3["Interactive Timelines & Embedded Video"] end UI --> G["Gemini Live with Regional Dialects"] ```

Woodward then unveils the app's first major pillar: a complete aesthetic and structural overhaul called “Neural Expressive.” Built from the ground up with fluid animations, custom typography, haptic cues, and vibrant colors, the redesign reimagines model outputs. Rather than serving static walls of text, Gemini generates bespoke, dynamic UI layouts in real time, featuring interactive timelines, expandable visual deep dives, and embedded video clips. Furthermore, Gemini Live now launches instantly inline and introduces localized voice persona dialects, including a distinct Liverpool accent.

- “Good design is about how it works,” Woodward noted, emphasizing that the redesign streamlines content remixing and visual generation templates. - The Neural Expressive redesign rolls out globally on Android, iOS, and the web immediately.

---

[01:13:01](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4381s) Creative Video Generation with Gemini Omni

The second major update integrates Google's flagship multimodal model, Gemini Omni, straight into the consumer Gemini application for paid subscribers. Woodward compares the current leap in generative video to the transformative inflection points seen previously in image generation, noting that users can manipulate raw footage using combinations of text prompts, image styles, and video inputs.

To demonstrate the system's spatial reasoning, Woodward introduces a workflow featuring a musical artist named Sasha creating a music video teaser. Sasha feeds a raw video clip of herself singing alongside reference visual art into the app. Gemini Omni transforms the aesthetic style of the scene and reorients the virtual camera into a sweeping 360-degree rotational shot. The model parses the physical dynamics of her movement and layers visual effects into the scene while maintaining facial fidelity and spatial coherence.

- Gemini Omni is available immediately within the Gemini app for Google AI+ Pro and Ultra tier subscribers globally. - “As you can see, Omni understands the physics of her movement,” Woodward observed. “It intelligently layers those effects into the real world without losing the soul of her shot.”

---

[01:14:48](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4488s) Proactive Assistance: The Daily Brief Agent

The third announcement pivots to proactive autonomous agency with the introduction of an out-of-the-box agent called the Daily Brief. Woodward stresses that AI agents represent a fundamental evolution beyond reactive chat prompts: “This is a big shift for Gemini because agents don’t just answer questions, they proactively work on your behalf.”

The Daily Brief functions as an automated morning digest that synthesizes pending tasks, incoming calendar appointments, and relevant emails across a user's Google ecosystem. Rather than delivering a raw summary, the agent clusters tasks by subject, surfaces easily forgotten reminders—such as an overdue library book—and suggests direct subsequent steps. For travel-related itineraries, the Daily Brief provides inline action cards allowing users to resolve confirmations or adjust reservations without navigating away from the briefing feed.

- The Daily Brief begins rolling out immediately to Google AI+ Pro and Ultra subscribers in the United States.

[01:15:55](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4555s) Gemini Spark Progress and Multi-Device Workflows

Returning to the custom workflows kicked off earlier via Gemini Spark on a mobile phone, the presenter transitions to a laptop to demonstrate seamless cross-device synchronization. The laptop screen shows that Spark has successfully broken down the initial request into three distinct subtasks labeled 1, 2, and 3. The system incorporates an intentional human-in-the-loop review mechanism: for sensitive actions, the agent pauses to request direct user approval rather than executing assumptions automatically.

Opening up the first completed task—a school year planning checklist designed to capture everything that needs to get done between now and the end of the year—the presenter reveals how Spark leverages native Google Docs capabilities. Rather than generating plain text, the output is fully structured: it organizes activities across various children, includes precise dates, and applies distinct color coding and formatting options natively supported by Google Docs, turning sprawling family logistics into an actionable, easily navigable checklist.

```mermaid flowchart TD accTitle: Gemini Spark Multi-Device Workflow accDescr: Diagram illustrating how Gemini Spark executes user requests across mobile and desktop environments with human approval checkpoints.

subgraph Input ["Task Creation"] A["User Request on Phone"] end

subgraph Sync ["Agent Processing & Sync"] B["Gemini Spark Core Engine"] C["Subdivided Tasks (1, 2, 3)"] end

subgraph Interaction ["Laptop Workspace"] D{"Requires Approval?"} E["User Approves Action"] F["Formatted Google Doc Output"] end

A --> B B --> C C --> D D -- Yes --> E D -- No --> F E --> F ```

- Spark rolls out to trusted testers this week. - A beta release launches for Google AI Ultra subscribers in the United States next week. - Workspace enterprise environments and Gemini Workspace are scheduled to receive dedicated Spark versions. - A future proactive capability highlighted will allow Spark to read upcoming calendar events (such as weekend tee-ball snack duty) and automatically initiate an Instacart order, remembering specific family constraints like allergy-conscious, nut-free snacks.

---

[01:18:22](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4702s) Gemini for macOS and Finder Dictation Demo

Building on the previous month's release of the native macOS Gemini app—which was built from scratch using Antigravity by a lean engineering team delivering over 100 features in under 100 days—the speaker presents an unannounced feature developed just two weeks prior. The use case addresses the tedious paperwork required when boarding pets at a new kennel, which traditionally demands hunting down medical documents, vaccine certificates, and allergy logs.

On a MacBook, the presenter selects multiple pet document files directly inside Finder—including invoice images and vaccination PDFs for their two dogs, Hank and "Louis Cinnamon." Holding down the keyboard's Function key, the presenter dictates an informal, conversational draft email requesting a short stay starting Thursday, before immediately correcting themselves on the fly: *"Oh, wait, no, actually, it's this Friday."* The speaker asks the model to extract the medical facts into an inline table, detail allergies, and ensure a friendly tone. Upon releasing the Function key, Gemini uses its multimodal understanding to parse the raw files in Finder, accurately incorporates the self-correction to Friday, and generates a structured email complete with an embedded table of dates, vaccines, and notes.

> “This is the power of what Gemini can do using your voice.”

- Capabilities demonstrated: Multimodal ingestion of Finder PDFs and images paired with real-time conversational voice dictation and correction handling. - Feature availability: These advanced voice dictation features and Gemini Spark integrations are arriving on the Gemini macOS app during the summer. - Keynote recap: The presentation highlights the Neural Expressive interface redesign, availability of the Gemini Omni model and 3.5 Flash, the upcoming 3.5 Pro model, and background automation via Daily Brief and Gemini Spark.

---

[01:21:49](https://www.youtube.com/watch?v=wYSncx9zLIU&t=4909s) Community Spotlight: Holly and Work Onward

The presentation shifts to an inspiring community story from Holly, an entrepreneur who emigrated from South Korea to the United States in 2005. Following her family's passion for culinary arts, she opened her own restaurant, navigating the everyday operational friction of inventory, bookkeeping, and scheduling. Holly turned to Gemini to streamline daily restaurant management—ranging from drafting marketing copy and planning menus to tracking supply orders via conversational queries (such as confirming five gallons of sesame oil ordered the previous month). Using Antigravity and Stitch, she built a website featuring an integrated customer-facing chatbot.

When an unexpected staffing shortage forced an urgent search for a dishwasher, Holly recognized a structural gap in existing employment boards for non-technical workers and independent proprietors. She assembled a team to develop **Work Onward**, an inclusive hiring platform powered by Gemini models. Work Onward removes language and technology barriers by translating listings automatically and enabling users to post job openings purely via standard SMS text messages.

> “It's about honoring the dignity of workers and small businesses, and I think that's the power of AI.”

- Scale: Over 13,000 workers and business owners across New York City actively utilize Work Onward. - Accessibility: Allows business owners and job seekers to interact entirely through plain text messaging without downloading dedicated software or mastering complex interfaces.

---

[01:23:39](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5019s) Generative Media: Google Pics and Stitch Updates

Suz Chambers takes the stage to discuss Google's generative media philosophy, emphasizing that technology serves as an enabler for human expression rather than an end in itself. Highlighting that products should shorten the latency between an initial creative spark and its execution, Chambers reveals three distinct creative tools, beginning with **Google Pics**, a new workspace utility that integrates generative image control with fine-grained editing capabilities.

Unlike blunt text-to-image prompts, Google Pics understands the distinct semantic objects inside a canvas. Users can begin with a base image, hover over individual visual elements to delete them, resize components to alter proportions, adjust framing, insert typography, and translate text within the composition with a single click. Every asset created or modified across Google’s generative tools, including Pics, is embedded with SynthID digital watermarking.

```mermaid flowchart LR accTitle: Stitch Design Flow accDescr: Diagram tracing the website UI creation process from a voice or text prompt to live production code using Stitch.

A["Idea & Concept"] -->|"Prompt / Voice"| B["Stitch Generation"] B -->|"Real-Time Tweaks"| C["Dynamic UI Canvas"] C -->|"Export"| D["Production Code / Live Website"] ```

Transitioning from static visual editing to software creation, Chambers introduces updates to **Stitch**, Google’s interface design generation system. Over the preceding year, creators and internal teams generated over 100 million UI screens with Stitch. Chambers demonstrates how small business owners—such as pizza parlor proprietors Tyler and Jenny—can vocalize adjustments in real time:

> “Now you can build UI at the speed of thought.”

Prompting or speaking commands such as *"make the header text larger and update the menu to highlight more pizza options"* triggers immediate, live canvas re-renders without manual coding. Once finalized, Stitch exports the visual layouts directly into clean code or publishes them to the web with minimal clicks, rolling out globally to users today. Chambers closes the section by pointing back to the launch of Google Flow at the prior year's I/O, previewing new creative collaborations across creative industries.

[01:28:14](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5294s) Generative Media: Gemini Omni and Creative Flow Tools

Google expands its generative media toolkit across video, audio, and custom workflows. With Gemini Omni, creators can transform video environments, insert custom visual effects, and add new characters from a simple text prompt and style reference without altering the actor's original cadence or body language. The presentation illustrates this with raw footage of a person walking, altering the surroundings and inserting assets while holding the subject’s pacing and presence completely steady.

```mermaid flowchart TD accTitle: Creative Flow Architecture accDescr: Pipeline showing how raw footage and images are transformed via Gemini Omni and Flow Agents into multimodal outputs.

subgraph Input ["User Inputs"] A["Raw Video / Single Image / Audio Riff"] B["Natural Language Prompt & Style Reference"] end

subgraph Engine ["Flow & Gemini Processing"] C["Gemini Omni Engine"] D["Flow Multi-Action Agent"] E["Flow Tools (Vibe Coding)"] end

subgraph Output ["Generated Artifacts"] F["Preserved Performance with New Characters/VFX"] G["Multi-Angle Generation (16 Unique Videos)"] H["Full Production Demos (Google Flow Music)"] end

A --> C B --> C C --> F A --> D D --> G B --> E E --> H ```

In Google Flow, a newly integrated multi-action agent replaces the older single-prompt limit, allowing users to orchestrate complex production steps in parallel. When given a single still image, the agent analyzes the scene geometry and composition, conceptualizes optimal perspectives, and automatically generates 16 unique video angles. The agent can also execute sweeping environmental transformations across an entire project—such as changing an early morning desert shot to midnight, adjusting lighting dynamically so headlights realistically illuminate kicked-up dust. Furthermore, Flow Tools introduces "vibe coding," enabling artists to build, share, and remix customized utilities for specific workflows, including custom video effects, hand-drawn animations, and typographic treatments.

On the audio front, Google Flow Music is designed to guide musicians through early track prototyping. To demonstrate, an engineer took a raw, unadorned piano riff and prompted Flow Music to generate an R&B track featuring guide female vocals to pitch to bandmates:

> “The real breakthrough isn't the technology, it's what you do with it.”

- **Omni Character Preservation**: Integrates new characters and scene styling while maintaining the source actor’s exact motion and timing. - **Multi-Angle Video Synthesis**: Converts a single source image into 16 distinct cinematic camera angles. - **Vibe Coding in Flow Tools**: Allows non-programmers to define bespoke image, animation, and text toolkits on the fly. - **Availability**: Gemini Omni features, Google Flow updates, and Google Flow Music are made available immediately.

---

[01:33:00](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5580s) Android XR and the Vision for Intelligent Eyewear

Shahram Izadi introduces the next operational milestone for Android XR, the platform developed alongside Samsung and optimized for Snapdragon silicon with Qualcomm. The platform splits hardware into two primary paradigms of wearable intelligence: display glasses with integrated micro-screens and lightweight audio glasses. Display glasses, previewed previously, provide glanceable visual data such as live translation, ride-share pickup details, and custom micro-widgets using "create my widget." An expanded trusted tester program for display eyewear is slated for later in the year.

The immediate commercial focus shifts to screenless audio glasses launching in fall. Instead of presenting visual overlays that pull a user's attention from their physical surroundings, these frames whisper context-aware Gemini assistance privately into the user's ear. This enables hands-free phone calls, image capture, audio streaming, and ambient app interaction while the paired smartphone stays stowed in a pocket. The devices are engineered for universal platform support, pairing natively with both Android and iOS devices.

- **Platform Foundations**: Built on Android XR in engineering partnership with Samsung, powered by Qualcomm Snapdragon platforms. - **Display Glasses Horizon**: In-lens heads-up information (such as Uber ride statuses and live translation) moving into an expanded trusted tester program later in the year. - **Audio Glasses Delivery**: Commercial debut scheduled for fall, delivering private, spoken Gemini interactions without visual distraction. - **Cross-Platform Compatibility**: Out-of-the-box pairing support for both Android and iOS smartphones.

---

[01:35:14](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5714s) Samsung, Warby Parker, and Gentle Monster Design Collaboration

Samsung's Jay Kim details the physical engineering behind the eyewear collection, explaining that consumer adoption hinges on treating glasses as personal fashion items rather than tech gadgets. Samsung handles the micron-scale engineering, sensor integration, and battery miniaturization, while working alongside design houses Warby Parker and Gentle Monster to produce distinct aesthetic lines. Gentle Monster focuses on bold, rebellious frame silhouettes that emphasize wearer confidence, whereas Warby Parker adapts classic, heritage eyewear styles that celebrate technical elements as honest design features rather than concealing them.

> “Intelligent eyewear that looks prettier than normal eyewear. That's our goal.”

The presentation highlights that wearable technology must prioritize personal expression, weight distribution, and subtle contours. Because frame fit directly impacts daily comfort, every millimeter of internal component layout is optimized to make the frames indistinguishable in balance from standard prescription glasses and luxury sunglasses.

---

[01:38:41](https://www.youtube.com/watch?v=wYSncx9zLIU&t=5921s) Live Demo: Ambient Navigation and Background App Actions

Shahram Izadi and Nishtha Bhatia step out to demonstrate the audio glasses live, with Izadi wearing a Warby Parker frame and Bhatia wearing a Gentle Monster design. For the live stage demonstration, the internal temple speakers are piped into the auditorium's main sound system so the audience can hear Gemini’s real-time conversational voice.

```mermaid flowchart LR accTitle: Background DoorDash Order Flow accDescr: Step-by-step sequence of Gemini triggering an app background task from ambient voice commands.

A["User Voice Command:<br/>'Order my usual'"] --> B["Gemini Audio Context Engine"] B --> C["Personal Intelligence / History Match"] C --> D["Background App Action:<br/>DoorDash Execution"] D --> E["Autonomous Menu Selection<br/>(Nitro Cold Brew)"] E --> F["Voice Confirmation & Tip Addition"] ```

The live interaction starts with ambient contextual navigation. Izadi triggers navigation using personal conversational history without naming a specific street or landmark:

> “Hey, can you navigate me to the place I met my friend Gianna at last week.”

Gemini parses his location history, identifies the Redwood Grove Nature Preserve, and suggests an en-route coffee stop based on his typical daily habits:

> “Want to add a stop on the way to grab your afternoon cold brew?”

When Izadi accepts, Gemini routes the walk via Coupa Cafe and gives directional orientation relative to the Bill Graham Parkway. Nishtha Bhatia then demonstrates multimodal mobile task execution without unlocking or pulling out her phone. While her device remains asleep in her pocket, she asks Gemini to place her usual order at that same cafe. A live mirror of her phone screen shows Gemini autonomously launching DoorDash, searching the Coupa Cafe catalog, selecting a Nitro Cold Brew, and advancing through the ordering screens to the final purchase page. Once Gemini asks for confirmation, Bhatia verbally authorizes the charge and adds a 20% tip completely hands-free.

- **Context-Aware Routing**: Resolves ambiguous locations (“where I met Gianna”) by synthesizing personal timeline history with Google Maps walking directions. - **Ambient Orientation**: Offers contextual landmarks and spatial cues, such as notifying the user that a shop is approaching on their right. - **Zero-Touch Background Tasking**: Gemini controls third-party applications (like DoorDash) in the background while the smartphone remains locked in a pocket. - **Conversational Confirmation**: Prompts the user verbally to authorize final payment amounts and adjust tipping percentages prior to placing orders.

[01:42:01](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6121s) Live Demo: Intelligent Eyewear, Multimodal Actions, and Wrist Previews

The live demonstration of the smart glasses continues as Nishtha catches up on missed notifications through voice interaction. Asking Gemini if any important updates were missed, the assistant responds that her family group chat arranged to meet for dinner that evening and reminds her to say her daily affirmations: “You are strong. You are confident.” Postponing the affirmations, she instructs Gemini to add the family dinner to her calendar. The assistant cross-references her existing commitments and responds: “Sure, I'll add an event for family dinner at 7:00 PM to your calendar right after your team celebration.” Gemini handles both message summarization and real-time calendar updates directly through voice.

```mermaid flowchart TD accTitle: Intelligent Eyewear Interaction Flow accDescr: Diagram tracing the live demonstration steps from the smart glasses voice prompt to watch display output.

subgraph AudioCalendar["Voice & Context Actions"] A["Voice Prompt via Glasses"] --> B["Gemini Assistant Processing"] B --> C["Summarizes Muted Group Chat"] B --> D["Schedules Family Dinner at 7:00 PM"] end

subgraph GenerativeCamera["Generative Image & Wearable Output"] E["Camera Capture of Audience"] --> F["Generative Restyle: Cartoon Blimp"] F --> G["Glanceable Preview on Smartwatch"] end

D -.-> E ```

The presentation moves to showing how intelligent eyewear interacts with secondary glanceable displays, such as smartwatches. For the traditional Google I/O audience selfie, a complex generative request is issued to the eyewear: “Gemini take a photo of this amazing audience, but turn it into a cartoon and add a big blimp in the sky that says Google I/O 2026 on it in fun colors.” While waiting for the processing, the presenter jokes that the audience should “go bananas” because “Nano banana sunglasses is just awesome.” Within moments, the stylized cartoon render appears seamlessly as a preview directly on her wrist-worn watch display.

- Hardware designs come from partner fashion and eyewear brands, combined with engineering and craftsmanship from Samsung. - Proactive intelligence is powered directly by Gemini alongside custom ecosystem applications from Google and third-party developers. - The first generation of intelligent glasses is slated for commercial arrival in the fall.

---

[01:44:37](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6277s) Demis Hassabis on AGI Horizons, Model Safety, and Code Mender

Demis Hassabis takes the stage to reflect on the day's key platform announcements—highlighting Gemini 3.5, Gemini Omni, Antigravity 2.0 coding workflows, Search agents, and Gemini Spark. He notes that artificial general intelligence is transitioning from a theoretical aspiration to an imminent milestone: “Even for those of us who have spent our entire lives working on this, AGI is now on the horizon and it will be the most profound and impactful technology ever invented.” He stresses that navigating this threshold requires balancing technological promise with rigorous systemic safety:

> “We're in a moment of immense promise, but also enormous responsibility. It's important that we are clear eyed about the potential challenges, and use all the tools at our disposal to ensure the safety of our agentic systems, and ultimately, AGI itself.”

Addressing one of the critical real-world risk vectors, cybersecurity, Hassabis details Google's work on automated vulnerability mitigation. Drawing upon decades of internal security research, Google developed Code Mender, an automated code security agent built to detect and autonomously repair critical software vulnerabilities across massive codebases. Hassabis announces that an initial select group of external security experts is being invited to test the new Code Mender API, with a broader public launch planned shortly after.

---

[01:46:17](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6377s) Gemini for Science, Alpha Earth Foundations, and Weather Next

Hassabis shares the personal motivation behind his career, stating that his primary goal in pursuing artificial intelligence has always been to build tools that advance science and our comprehension of the natural world. To accelerate empirical work, he introduces Gemini for Science, a specialized suite designed to automate time-consuming academic routines. New prototypes coming out of Google Labs assist researchers with tracking high volumes of newly published literature, translating high-level research targets into executable code, and generating novel scientific hypotheses.

```mermaid flowchart LR accTitle: Scientific AI Models and Applications accDescr: Overview of Google scientific AI engines and their application domains.

subgraph Workflows["Scientific Tooling"] A["Gemini for Science"] -->|"Hypothesis & Code Gen"| W1["Lab Research Tasks"] end

subgraph Simulation["Earth & Planetary Scale"] B["Alpha Earth Foundations"] -->|"Digital Twin"| S1["Deforestation & Food Security"] C["Weather Next"] -->|"Path & Intensity AI"| S2["Hurricane Early Warning"] end

subgraph Biological["Molecular Modeling"] D["AlphaFold & Alpha Genome"] -->|"Virtual Cell Simulation"| B1["Structural Biology"] E["Isomorphic Labs"] -->|"Molecular Interaction"| B2["Oncology & Immunology Pipelines"] end ```

Beyond text and code generation, Hassabis emphasizes simulation as an indispensable pillar for understanding complex dynamic systems that defy classical modeling: - **Alpha Earth Foundations**: Functions as an emergent digital twin of the Earth, helping researchers model environmental dynamics, track deforestation patterns, and support food security initiatives. - **Weather Next**: A state-of-the-art global weather forecasting model designed to predict severe storm paths and intensification faster and with greater spatial accuracy than conventional numerical modeling frameworks. - **Hurricane Season Validation**: A video case study highlights Weather Next's 2025 track record, during which it predicted a Category 5 hurricane striking Jamaica three days earlier than standard operational guidance. Forecasters at the Hurricane Center noted the advance warning gave authorities the lead time needed to issue targeted evacuation orders and protect lives during Hurricane Melissa.

---

[01:48:56](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6536s) Accelerating Biology, Isomorphic Labs, and the Singular Horizon

Looking toward future biological frontiers, Hassabis envisions modeling emergent biological behavior down to complete virtual cells. Existing foundational biology models like AlphaFold and Alpha Genome have matured into essential research instruments utilized by millions of scientists globally, establishing what Hassabis terms “science at digital speed”—dramatically accelerating both the discovery of solutions and their worldwide distribution.

He reaffirms his long-standing conviction regarding the ultimate purpose of artificial intelligence: “I've always believed the number one application of AI should be to improve human health.” Through Isomorphic Labs, molecular interactions are modeled computationally to radically compress drug development timelines alongside biopharma partners. Several internal programs have now advanced into the preclinical stage, focusing on novel therapeutics for complex immune disorders and oncology. Hassabis summarizes their ambition directly: “Our mission is to reimagine the drug discovery process with the goal of one day solving all disease, something that would have seemed impossible just a few years ago.”

In his closing remarks, Hassabis contextualizes the current era within the broader arc of human history, reflecting on the profound shift underway:

> “When we look back at this time, I think we all realize that we were standing in the foothills of the singularity. It will be a profound moment for humanity. This technology will be a force multiplier for human ingenuity and usher in a new golden age of scientific discovery and progress, improving the lives of everyone, everywhere.”

## Notable quotes > [01:44:37](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6277s) Even for those of us who have spent our entire lives working on this, AGI is now on the horizon and it will be the most profound and impactful technology ever invented. — Demis Hassabis

> [25:26](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1526s) We've moved beyond AI tools that help us write to agents that help us act. These agents have lowered the barrier to development so much that anyone can be a builder, even busy CEOs. — Varun Mohan

> [29:34](https://www.youtube.com/watch?v=wYSncx9zLIU&t=1774s) Multi-day engineering efforts are collapsing into hours, if not minutes. — Varun Mohan

> [01:48:56](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6536s) I've always believed the number one application of AI should be to improve human health. — Demis Hassabis

> [01:48:56](https://www.youtube.com/watch?v=wYSncx9zLIU&t=6536s) Our mission is to reimagine the drug discovery process with the goal of one day solving all disease, something that would have seemed impossible just a few years ago. — Demis Hassabis

## Mentioned Google, Sundar Pichai, Demis Hassabis, Varun Mohan, Josh Woodward, Liz Reid, Vidhya Srinivasan, Suz Chambers, Shahram Izadi, Nishtha Bhatia, Jay Kim, Holly, Gemini, Gemini 3.5 Flash, Gemini 3.5 Pro, Gemini 3.1 Pro, Gemini Omni, Gemini Omni Flash, Gemini Omni Pro, Gemini Spark, Gemini Live, Google Antigravity, Antigravity 2.0, Google Search, AI Overviews, AI Mode, TPU, TPU v8 Training, TPU v8 Inference, SynthID, Content Credentials, C2PA, Circle to Search, Google Chrome, Android Studio, Firebase, Google AI Studio, Universal Commerce Protocol (UCP), Agent Payments Protocol (AP2), Universal Cart, Google Wallet, Google Pay, NotebookLM, Daily Brief, Google Docs, Google Sheets, Google Slides, Google Drive, Gmail, Google Keep, Google Calendar, Google Photos, YouTube, Ask YouTube, Google Maps, Stitch, Google Pics, Google Flow, Flow Tools, Google Flow Music, Android XR, Samsung, Qualcomm, Snapdragon, Warby Parker, Gentle Monster, Code Mender, Gemini for Science, Alpha Earth Foundations, Weather Next, AlphaFold, Alpha Genome, Isomorphic Labs, OpenAI, Meta, Microsoft, Amazon, Salesforce, Stripe, NVIDIA, ElevenLabs, Kakao, DoorDash, Work Onward, Model Context Protocol

## Video description It's time to I/O! Tune in to learn the latest news, announcements, and AI updates from Google.

00:00 Opening 17:10 Gemini Omni 24:43 Google Antigravity 35:13 Gemini Spark 43:32 Search 01:07:52 Gemini 01:23:26 Generative Media 01:32:47 Intelligent Eyewear 01:44:28 Closing

To watch this keynote interpreted in American Sign Language (ASL), please click here: https://youtube.com/watch?v=vw73r_xaeX0

#GoogleIO

Subscribe to our Channel: https://www.youtube.com/google Find us on X: https://twitter.com/google Watch us on TikTok: https://tiktok.com/@google Follow us on Instagram: https://www.instagram.com/google Join us on Facebook: https://www.facebook.com/Google

[Source](https://www.youtube.com/watch?v=wYSncx9zLIU)

Reading summary

At Google I/O '26, CEO Sundar Pichai and leadership framed Google's operations around an "agentic era," demonstrating autonomous software development, cross-platform background task agents, and an industry-wide commerce protocol. Underpinning these consumer and enterprise tools is a massive hardware expansion, with projected annual capital expenditures reaching up to $190 billion. Capital Expenditure and Dual-Chip TPU Infrastructure (11:17) To sustain global AI demand, Google projected its annual capital expenditure to reach between $180 billion and $190 billion, roughly a sixfold increase from $31 billion in 2022. Marking a decade of custom silicon development, the company unveiled its eighth-generation Tensor Processing Units, dividing its microarchitecture into dedicated chips for training (TPU v8 Training) and low-latency serving (TPU v8 Inference). According to Google, these clusters operate across distributed data centers orchestrated via JAX and Pathways, delivering up to twice the performance per watt. Autonomous Development via Antigravity 2.0 (25:26) Google detailed a transition from code-completion tools to autonomous acting agents anchored by Gemini 3.5 Flash and Antigravity 2.0, an agent-first desktop application equipped with a dedicated CLI and SDK. In an internal benchmark test, the Antigravity harness coordinated parallel subagents to generate an operational operating system from scratch—including its scheduler, memory subsystem, and virtual file system—in 12 hours for less than $1,000. Agentic Search and Asynchronous Productivity Workflows (43:42) Google reported that Search AI Mode exceeded 1 billion monthly active users, prompting a full unification of AI Overviews and conversational search. Rather than displaying static search listings, Search now compiles dynamic, interactive user interfaces and multi-session stateful planning tools directly on the fly using Antigravity sandboxes. Liz Reid announced continuous background "information agents" launching in summer to track complex criteria such as real-time financial metrics, rental listings, and product drops. In parallel, Google introduced Gemini Spark, a background engine operating on cloud virtual machines to deconstruct and execute multi-step user prompts across Gmail, Docs, Sheets, and Calendar. Demonstrated workflows included tracking event RSVPs and compiling team update emails through customized behavioral style guides. Spark incorporates human-in-the-loop checkpoints for sensitive actions and plans to expand natively into Chrome and an Android Halo hub later in the year. Open Frameworks for Autonomous Agent Commerce (58:45) Addressing consumer shopping, Google announced three commercial standards built on its 60-billion-item Shopping Graph. The Universal Commerce Protocol (UCP), an open-source standard co-developed with partners including Amazon, Meta, Microsoft, Salesforce, and Stripe, coordinates product discovery, checkout, and tracking across disparate merchant platforms. To mitigate security concerns regarding unauthorized agent purchases, the Agent Payments Protocol (AP2) enforces user-defined budgetary limits and cryptographically verifiable digital mandates. These protocols integrate into the Universal Cart, which tracks price histories and validates technical component compatibility across Google surfaces. Ambient Wearables and Frontier Scientific Research (93:00) In hardware, Google introduced Android XR-powered audio glasses developed in partnership with Samsung, Qualcomm, Warby Parker, and Gentle Monster. Scheduled for a fall release, the screenless eyewear delivers ambient, contextual Gemini voice assistance, handles hands-free background app tasks like food ordering, and syncs glanceable previews to smartwatches, while display glasses remain in trusted testing. Demis Hassabis forecast that artificial general intelligence is "just a few years away," arguing that systemic safety must balance rapid capability leaps. To address software security vulnerabilities, Google introduced the Code Mender API for automated vulnerability remediation. In scientific applications, Google highlighted Gemini for Science, climate modeling via Weather Next, and biological simulations through Isomorphic Labs, which has progressed computational drug discovery pipelines into preclinical stages for oncology and immune disorders. Google's keynote presented an aggressive full-stack consolidation, migrating its search, operating systems, and consumer apps from passive query tools into autonomous background agents. However, the commercial viability of this ecosystem depends heavily on consumer trust in autonomous financial and data transactions, developer adoption of its commerce protocols, and the return on escalating capital expenditures.

Source quotes

[104:37] Demis Hassabis: Even for those of us who have spent our entire lives working on this, AGI is now on the horizon and it will be the most profound and impactful technology ever invented. [25:26] Varun Mohan: We've moved beyond AI tools that help us write to agents that help us act. These agents have lowered the barrier to development so much that anyone can be a builder, even busy CEOs. [29:34] Varun Mohan: Multi-day engineering efforts are collapsing into hours, if not minutes. [108:56] Demis Hassabis: I've always believed the number one application of AI should be to improve human health. [108:56] Demis Hassabis: Our mission is to reimagine the drug discovery process with the goal of one day solving all disease, something that would have seemed impossible just a few years ago.

Original transcript

[0:01] Take a second to realize how far we've come. [0:07] Empowered by all the things we've made together. [0:10] How many people have looked at a problem [0:12] and then been bold enough to say, I can solve it. [0:15] I'm staring at a future we all get to decide. [0:19] As the ultimate tool to solve all the world's [0:22] most complex scientific problems. [0:24] We're taking flight. [0:29] Every one of us is packed with the potential [0:35] to make anything we can dream of. [0:38] I do not know personally how to code. [0:40] But given the option to create something just [0:43] opens up so many doors for me as an educator. [0:47] I discovered Gemini could help me with a lot of things, [0:50] from menus to marketing and inventory. [0:54] But in a world where anything is possible, [0:56] we're really pushing at this big mission. [0:58] The mission of solvable disease. [1:00] This is huge, but we have a profoundly new technology [1:02] in our hands. [1:03] What matters is what we choose to build. [1:07] Hi, my name is Matthias. [1:09] My brother Tommaso made me in this AI assistant. [1:13] So let's not just make. [1:21] Let's make something that matters. [1:54] SUNDAR PICHAI: Hello, Shoreline, and hello to everyone [1:57] watching from around the world. [2:00] Excited to be back for this year's I/O. [2:03] I think it's been an incredible year. [2:06] All of the relentless shipping, the rapid advances [2:09] in technology. [2:11] It's been a period of progress. [2:14] I've definitely felt it myself. [2:16] It's been an intense year. [2:18] Here's a look at what I've been up to. [2:26] OK I wish this was my year. [2:29] Actually, one of me plugging in the TPU is pretty accurate. [2:35] So there's still more work to do before it's in space. [2:37] We are working on that. [2:40] On a serious note, it's been an extraordinary moment. [2:44] It's been 10 years since we pivoted the company [2:46] to be AI first. [2:48] We knew how profound I would be to advancing our mission [2:53] and improving people's lives at scale. [2:56] This is why we are taking a differentiated, full stack [3:00] approach to AI innovation. [3:02] From our custom silicon and secure foundation, [3:06] to our world class research and models, [3:09] to our products and platforms that reach billions of people. [3:13] This approach enables us to iterate and innovate faster, [3:17] and it's lighting up every part of the company. [3:20] What's really incredible is how people are using our AI. [3:23] Students prepping for final exams with the Gemini app. [3:27] Musicians and artists using generative AI models [3:30] like Lyria and Veo as part of their creative flow. [3:34] Developers coding and bringing their ideas to life. [3:38] I've been using Gemini in a myriad of ways. [3:40] Recently, I've been turning to Gemini [3:42] to make sense of my parents doctor visits. [3:45] I'm sure many of you have done a version of that. [3:48] These stories of how people are using AI [3:50] are the best measure of progress to understand the scale at which [3:55] people are adopting AI. [3:56] There's another great proxy tokens. [3:59] The fundamental units of data are models process, [4:02] many representing a problem being solved. [4:05] Two years, we were processing 9.7 trillion tokens a month [4:09] across the surfaces, a huge number. [4:12] Last year at I/O that grew to about 480 trillion tokens. [4:17] And fast forward to today, that number [4:19] has jumped seven times to 3.2 quadrillion tokens per month. [4:26] I never imagined I'd say quadrillion in an I/O keynote. [4:33] But here we are. [4:34] Now, some out there might call this token maxing. [4:37] And there's probably some truth to it. [4:41] I still think it tells an important story [4:44] about our products and how others are building as well. [4:48] Especially our developers. [4:50] Purse over 8.5 million of you are now [4:53] building new apps and experiences with our models [4:56] monthly, and our model APIs are now [4:59] processing around $19 billion tokens permanent. [5:03] Over the past 12 months, over 375 customers [5:08] each processed more than one trillion tokens, [5:11] representing incredible demand for AI across the industry. [5:16] We are, of course, also seeing incredible demand [5:18] across our products. [5:20] We now have 13 products with over a billion users each. [5:24] Five of those have more than three billion users. [5:34] Our Gemini models are a big reason [5:36] more people are using our products, [5:38] and why they are using our products more. [5:41] It all starts with search, which is bringing [5:43] the benefits of generative AI to more [5:45] people than any other product in the world. [5:48] AI Overviews now has over 2.5 billion monthly users, [5:52] and AI Mode has been a revelation. [5:55] Our biggest upgrade to search ever. [5:57] People love it. [5:58] In just a year, it's already surpassed [6:00] 1 billion monthly users. [6:02] When people use our AI powered features in search, [6:05] they use search more. [6:07] I love how search has become less about individual queries [6:11] and feels more like an ongoing conversation, [6:14] giving you deeper insights and connecting you [6:17] with the vastness of the web. [6:19] Another place we've been rapidly innovating is in the Gemini app. [6:23] Last year at I/O the Gemini app had 400 million [6:26] monthly active users. [6:27] Today we have surpassed 900 million. [6:29] More than doubling in a year. [6:37] In that same time, daily requests [6:39] have grown over 7 times. [6:41] It's incredible growth. [6:42] We've been adding a lot of unique features [6:44] like personal intelligence, which makes responses [6:47] more customized and helpful. [6:49] And to date, more than 50 billion images [6:52] have been generated with our Nano banana models. [6:55] It was a breakout star this past year. [6:57] I know you all have been having a lot of fun with it. [7:01] Beyond the Gemini app, we are also [7:03] having much more natural conversations [7:05] with Gemini directly inside many of our products. [7:09] Recently, Maps got its biggest upgrade in a decade, [7:13] including a new feature called Maps. [7:16] People are using it to ask more complex and much longer [7:20] questions. [7:21] Here's a real query from a parent. [7:24] My kid just fell into the duck pond [7:26] and the wedding starts in 30 minutes. [7:29] Where can I walk and buy her a new dress. [7:33] I'd like to hear how that turned out. [7:36] We're also bringing this conversational AI [7:38] to two more products. [7:40] First, ask YouTube. [7:42] People come to YouTube every day to ask a lot of questions. [7:46] There's a lot of great videos. [7:48] Sometimes it's hard to know where to start. [7:51] Ask YouTube entirely reimagines the experience. [7:55] Say you want to teach your three-year-old how [7:57] to ride a pedal bike, and they already [7:59] know how to ride a balance bike. [8:01] Just ask YouTube. [8:03] You'll see a couple of differences in results. [8:06] The information is digestible and easy to navigate. [8:10] You get an overview and helpful tips. [8:13] You will see videos that best match your interest. [8:17] So if you want to try a specific method of teaching [8:19] can go deeper there. [8:21] And best of all, it jumps right to the part of the video most [8:24] relevant for you. [8:26] Brings back memories of teaching my kids to ride. [8:30] It remembers the context so you can follow up with questions [8:33] like, should I buy one with handbrakes or pedal brakes. [8:36] Making it an ongoing conversation. [8:39] It even lays out the information in a table. [8:41] So it's easy to compare. [8:44] We are starting to test ask YouTube now [8:46] and it will roll out broadly in the US this summer so far. [8:57] So far, we have shown conversational text queries. [9:01] There are a lot of times I want to get things [9:02] done at the speed of my voice. [9:05] There is much more possible today [9:06] thanks to technical leaps in our audio models. [9:09] A new feature called docs live takes this to another level. [9:14] To create a doc with Gemini app, before you'd [9:16] have to type up a really precise prompt. [9:19] With docs live, you can just verbally [9:21] brain dump whatever is on your mind and let Gemini do the rest. [9:26] Let's see it in action with a demo from our product team. [9:29] This is all in real time, not sped up. [9:35] All right, let's try this out. [9:37] So I just remembered I'm doing an alumni talk for my High [9:40] School's career day tomorrow. [9:42] I need to come up with some talking points [9:44] to explain what I do for a living as a software engineer, [9:46] but I'm not really sure where to start. [9:49] Oh, actually, can you just pull my resume from drive. [9:53] Although that might be boring. [9:56] Maybe can you come up with some funny analogies [9:58] so it'll be more of an engaging talk for the students. [10:01] Oh, and also, I think the school sent me an email. [10:05] I think the subject is something like career day logistics. [10:08] Maybe just grab the details from there, [10:10] throw them at the top of the doc so I know where to go [10:12] and what time to get there. [10:14] Let's just update those requirements [10:15] and turn this straight into a draft. [10:26] This is cool, but it's a little dense. [10:28] Maybe format the analogies as a table. [10:30] So it's a little easier for me to scan and also add a note [10:33] to tell the story about how my brother inspired me to become [10:36] a software engineer, at the top of my doc and bold it [10:39] so I don't miss it. [10:48] Yeah, it looks great. [10:57] In the future, you'll be able to create new docs [11:00] and edit them directly, all with your voice. [11:03] Docs live is rolling out for Pro and Ultra subscribers [11:06] this summer, and the same powerful voice capabilities [11:09] will come to Gmail and Google Keep then too. [11:17] Incredible to see the pace of innovation rolling out [11:20] across our products, supporting all of this at scale [11:24] for our users, while also serving enterprises [11:27] and developers around the world, requires massive investments [11:30] in infrastructure, and we've been investing [11:33] for today and for the future. [11:35] In 2022, we were spending $31 billion annually in CapEx. [11:40] This year. [11:41] We expect that number to be about six times [11:43] that, approximately 180 to $190 billion. [11:49] A key part of this investment is our custom silicon. [11:53] A decade ago, we announced our very first commercial Tensor [11:56] processing unit, or TPU. [11:58] On this IO stage. [12:00] Since then, we have transformed how the industry builds for AI. [12:04] We recently announced our eighth generation of TPUs [12:07] at Cloud Next. [12:09] For the first time, we have taken a dual chip approach [12:11] with specialized architectures for training and inference TPU [12:15] 80 and 80. [12:17] While they may look similar, they're [12:19] actually pretty different. [12:20] 80 is optimized for large scale pre-training, [12:24] and it's nearly three times the raw computing power [12:26] of our previous generation. [12:29] We have taken a fundamentally different approach [12:31] with our training infrastructure, [12:33] with Jackson pathways. [12:34] Our training is no longer constrained [12:37] by the limits of a single massive data center. [12:40] Instead, we can now seamlessly distribute training [12:43] across multiple sites, scaling across more than TPUs globally. [12:48] This gives us the ability to create the largest training [12:51] cluster in the world for model builders. [12:55] This means training larger, more capable models in weeks [12:58] rather than months. [13:00] TPU is designed for inference. [13:03] We have dramatically improved speed at every step [13:07] because if we learn anything in 27 years of working on search, [13:10] it's that latency matters. [13:12] To give you a live sense of what this speed feels like, [13:16] here's a prompt on an upcoming Flash model. [13:19] If it were running on 80, I'll ask it [13:22] to create a Chrome Dino game. [13:24] Push submit. [13:25] The response is generated in real time as you watch. [13:30] Take a look at the tokens per second in the top right corner. [13:34] The speed is pretty incredible, nearly 1,500 tokens per second. [13:40] It almost took longer to write out the request. [13:43] And the game is pretty fun too. [13:52] In addition to speed, we are also [13:53] thinking about scaling sustainably. [13:55] Both chips are more energy efficient, [13:57] delivering up to two times better performance per Watt. [14:02] TPUs have been hard at work training for I/O this year. [14:06] I'm told we have a behind the scenes look. [14:21] Hey good weekend. [14:22] Yeah just folded proteins across rare Oncology Data sets. [14:26] You simulated the next 50 years of climate data. [14:28] I made a picture of a pug. [14:30] You ever see a pug dressed like an accountant. [14:34] No you want to. [14:38] All right, listen up. [14:39] TPUs IO starting soon. [14:42] And we got a job to do. [14:43] Actually, we got trillions of jobs to do. [14:45] So clear cache. [14:46] Timmy dry your fans. [14:49] Let's find out. [15:06] Past the point of no return. [15:09] Reach the top. [15:10] But still gotta learn how to keep it. [15:13] Hey, what are you doing. [15:14] I'm doing the montage thing. [15:16] Yeah, well, how about you montage your way down here [15:18] and help out. [15:19] Now All right. [15:22] Coming welcome to the limit. [15:25] Hey, don't look down. [15:28] Just keep your head up. [15:29] You'll be fine. [15:38] I bet Timmy TPU will be ready to teraflop right into bed [15:42] after IO. [15:44] Our compute innovations enable our advances. [15:48] There are three areas where I want [15:49] to go deeper today to show you the progress in each model's [15:53] coding and agents. [15:56] Let's start with the exciting progress in World models. [15:59] With world models, AI is moving from predicting text [16:04] to simulating reality. [16:06] Demis and the team at Google DeepMind [16:08] have been working to push the boundaries of what [16:11] these models can do. [16:12] Let me invite Demis out to share more. [16:28] DEMIS HASSABIS: Hi, everyone. [16:28] It's really great to be here. [16:30] Over the past year, AI capabilities [16:33] have leaped forwards. [16:34] We now have agents that can plan and act on our behalf. [16:38] And artificial general intelligence [16:40] is just a few years away. [16:42] Today, I'm excited to share the progress we've [16:45] made towards building AGI. [16:47] Last year, I outlined our vision of extending [16:50] Gemini's incredible multimodal capabilities [16:53] to become a world model AI that can understand [16:57] and simulate the world. [17:00] This is a crucial aspect of achieving AGI, [17:02] and we will be important for everything [17:04] from building AI assistants to training robots. [17:09] Now we're taking the next big step. [17:11] I'm excited to announce Gemini Omni. [17:19] Our new model that can create anything from any input. [17:24] It combines Gemini app intelligence [17:26] with the best of our generative media models [17:28] for a new level of World understanding multimodality [17:32] and editing models like Veo, Nano, banana, and Genie [17:36] are able to create extremely realistic videos, images, [17:40] and interactive simulations. [17:42] Although not perfect, they already [17:44] demonstrate some impressive notions of intuitive physics, [17:48] and with Omni, we've now made even more progress. [17:51] It's a step change in simulating things like kinetic energy [17:55] and gravity. [17:56] Previous systems would have found these concepts difficult. [18:00] Gemini App World knowledge and reasoning really shine in Omni. [18:04] It can translate complex ideas into highly accurate videos. [18:09] So for example, you can give it a simple prompt [18:11] make a claymation explainer of protein folding [18:14] and get this protein's start. [18:16] As chains of amino acids, they fold [18:19] into patterns like the alpha helix [18:21] and sections called beta Sheets, forming [18:24] a perfect 3 dimensional shape, but the initial generation [18:28] is just the start. [18:29] The creative process is rarely a single step. [18:32] It's usually iterative, just like Nano banana. [18:35] Redefine image editing. [18:37] Omni gives you a more natural way [18:38] to edit video with conversational language. [18:42] What's really cool is you can give it your own videos. [18:45] For example, this selfie and change [18:48] reality in a really fun way. [18:53] You can easily adjust the details and style [18:56] or even add elements, and the whole scene morphed in [19:00] to reflect your new idea. [19:08] A simple circle turns into a black hole, [19:10] or an evening stroll comes to life. [19:13] Anything becomes a Canvas for creating entirely new realities. [19:17] Let's take a look at what Omni can do. [19:22] I think you can. [19:45] See it starting with video. [20:05] But over time, Omni will be able to generate [20:07] any output from any input. [20:10] This was always our goal with Gemini [20:12] and why we built it to be multimodal from the very start. [20:16] It was a harder path, but the foundation is now paying off. [20:20] Today we're launching the first model in the Omni family Gemini [20:24] Omni Flash. [20:25] It's now available across our products [20:28] and you'll hear more about this later. [20:30] We're excited with the progress we're making, [20:32] and we'll be able to share more about omnipro soon. [20:36] We can't wait to see what you create. [20:38] Back to you, Sundar. [20:47] SUNDAR PICHAI: Thanks, Demis. [20:48] This is huge progress. [20:50] As generative AI gets better, so does the need [20:53] for greater transparency. [20:55] Research shows people can correctly [20:57] identify high quality deepfake videos only [21:01] about a quarter of the time. [21:03] Three years ago, we launched synthetic, our watermark that [21:06] is invisible to the naked eye. [21:09] Since launch, synthetic has now watermarked over 100 billion [21:13] images and videos, along with 60,000 years of audio assets. [21:18] Millions of people are using our SynthID detector [21:20] in the Gemini app to verify AI generated content. [21:25] We are now going a step further and adding content credentials [21:29] verification across products. [21:32] This will show if the origin of the content was AI or a camera, [21:37] and if it's been edited with generative AI tools. [21:40] In this example, Gemini can tell this photo was [21:44] captured with the Pixel camera and then edited [21:47] with Google Photos. [21:49] We want more people to have easy access to these tools, [21:53] so we are expanding both SynthID and content credentials [21:56] verification to search and Chrome. [22:05] You can simply circle to search or right click in Chrome [22:08] and ask, was this generated with AI. [22:11] And you'll get a clear response along [22:13] with other helpful context. [22:15] For example, this image was making the rounds [22:18] on social media last year. [22:21] It's obviously fake. [22:23] I don't eat hamburgers. [22:26] It might not be as clear to everyone else. [22:28] That's where these tools can be really useful. [22:32] Of course, this only works at scale. [22:34] If more partners decide to watermark their own AI [22:37] generated content. [22:39] NVIDIA signed on to SynthID last year, [22:42] and today I'm thrilled to announce [22:44] that OpenAI, Kakao, and Levin labs are adopting SynthID two. [22:55] It's great to see the cross-industry collaboration. [22:58] We are looking forward to expanding to more partners [23:01] and setting the standard of transparency for the AI era. [23:05] That's a look at the progress we are making with world models. [23:09] Now let's talk about what's next for our Gemini 3 family. [23:13] Gemini 3, launched a few months ago [23:15] with a full family of models. [23:17] It's been our most adopted series yet. [23:20] We have loved seeing developers use Flash as their daily driver [23:24] and build incredible experiences with pros, deep reasoning [23:27] and multimodal capabilities. [23:30] We've been hard at work on improving [23:31] these models, especially focused on agentic coding. [23:36] Long horizon tasks and real world workflows. [23:40] And today, I'm excited to introduce [23:42] Gemini 3.5 Flash, our first in a series of models. [23:50] Combining frontier intelligence with action. [23:52] Two things I would highlight. [23:54] First, when compared to 3.1 Pro Flash [23:58] is better across the board. [24:00] Almost all benchmarks. [24:02] It's made huge progress in coding. [24:05] And look at that extraordinary jump in GDP. [24:07] Val a benchmark that captures many real world, economically [24:11] valuable tasks. [24:13] Second, 3.5 flashes a very capable model at the frontier [24:18] and comparable to the best models, but much, much faster. [24:23] You look at the intelligence versus output speed, [24:26] it's in a whole league of its own in the top right quadrant. [24:30] When looking at output tokens per second, [24:33] it's four times faster than other frontier models, [24:36] and it's incredible delight to use. [24:39] The new model has been a game changer for us [24:41] internally at Google. [24:43] We've been using 3.5 Flash with the reimagined version [24:47] of our agent first development platform, Antigravity, [24:50] and it's dramatically accelerated how we build. [24:53] In March, we were processing half a trillion tokens a day [24:57] internally for our developers. [24:59] We've been doubling every few weeks, [25:02] and now we are processing more than three trillion tokens [25:05] a day. [25:06] This scale has created a powerful feedback loop [25:10] helping us improve. [25:11] 3.5 and of course, we are bringing it today [25:14] to developers in Antigravity. [25:16] Varun is going to share more. [25:26] VARUN MOHAN: It's truly an amazing time to be a builder. [25:29] We've moved beyond AI tools that help us write [25:31] to agents that help us act. [25:34] These agents have lowered the barrier to development [25:37] so much that anyone can be a builder, even busy CEOs. [25:41] In fact, Sundar used Google Antigravity last week [25:44] to fix a bug in the Google code base. [25:49] When we launched the Antigravity ID in November, [25:51] we made sure to nail the core agent powered ID experience [25:55] and added an experimental, first of its kind agent for surface [25:59] as a glimpse of where we were heading. [26:02] Millions of you are already actively use Antigravity, [26:05] and so we're excited to bring you even more today. [26:08] We've seen the diversity of tasks, preferences and frankly, [26:11] product feedback. [26:13] We've taken all these learnings, and now Antigravity is massively [26:17] expanding its suite of agentic capabilities, surfaces, [26:21] integrations and product features. [26:24] To start, we're launching a full CLI experience and Antigravity [26:29] SDK native voice support with Gemini audio models [26:33] and integrations with many services and platforms [26:36] like Android, Firebase and Google AI Studio. [26:47] All of this is available for you to try today. [26:51] But most importantly, at the core [26:53] is Antigravity 2.0, a new standalone desktop application [26:57] that delivers fully on that original glimpse of a truly [27:01] agent optimized experience. [27:03] The new Antigravity is unabashedly agent-first [27:07] focusing on the core agent conversations, agent produced [27:11] artifacts, and multi-agent orchestration. [27:14] I said, unabashedly agent first. [27:18] As Sundar mentioned, this is the exact experience [27:20] teams here at Google have been using to drive massive value. [27:24] The Antigravity agent harness, the invisible framework [27:27] for Gemini to perform real world tasks has become much more [27:30] powerful, with new core primitives [27:33] such as subagents, hooks, and asynchronous task management. [27:37] And underpinning all of this are the Gemini models, [27:41] with Gemini 3.5 Flash having been optimized [27:44] with the Antigravity harness. [27:47] Of course, being engineers, we were [27:50] curious to see how far we could push [27:51] the limits of what was possible with these agents and models. [27:55] So using the new Antigravity and Gemini 3.5 Flash, [27:59] we asked our agents to take on what [28:00] we consider to be a highly complex and impressive task. [28:04] Build a working operating system from scratch. [28:08] We were surprised by what we found asynchronously. [28:12] Antigravity broke down the challenge into a cohesive plan, [28:15] tackled tasks via parallel subagents generated, executed, [28:19] and iterated over its very own tests over 12 hours, [28:24] 93 subagents, working in parallel, [28:26] made over 15,000 model requests and processed 2.6 billion tokens [28:32] to take an initially empty project to the core [28:35] of a functioning operating system. [28:38] This was not possible. [28:46] This was not possible on Gemini 3.1 Pro, [28:50] but thanks to the performance and cost efficiency of Gemini [28:53] 3.5 Flash building, an entirely functional operating system [28:57] consumed less than $1,000 of API credits. [29:08] The Antigravity agents wrote every line of code, [29:11] from the scheduler to the memory management to the file [29:14] system generated, audited, and tested entirely by an autonomous [29:19] team of agents. [29:20] To put this in context, developing an OS from scratch [29:24] is both notoriously brutal and can take many months to build. [29:28] We weren't just building an application, [29:30] but a fully functioning operating system [29:32] that applications can run on. [29:34] Let's take this live and actually show this operating [29:36] system in action. [29:39] So here I'm actually in a terminal window in the OS [29:42] that Antigravity built. Now it's not super easy to demo a working [29:45] operating system. [29:46] So let's try something fun to see if it works. [29:49] One interesting utility that you can install [29:51] is SL, a common typo for RLS without spoiling it. [29:55] Here it goes. [29:58] It works. [30:01] You could see a cool locomotive passing [30:03] through the screen with the Antigravity logo on it, [30:06] but clearly this isn't a real OS unless I can play doom. [30:14] Now, if I try running doom right now, it just doesn't work. [30:18] Turns out that the OS is currently [30:19] missing some necessary video and keyboard drivers, so let's [30:23] just try and fix it in the new Antigravity. [30:26] I have a prompt prepared and I'm going to paste it in. [30:30] While it's running. [30:31] Let's take a tour of Antigravity 2.0. [30:33] As you can see, Antigravity 2.0 is very much agent. [30:37] First, with all of your conversations on the left side [30:40] panel and all of your projects. [30:41] Let's take a peek at one of the conversations [30:43] that I previously had. [30:46] I was curious for this demo about some fun facts about doom, [30:49] so I asked the agent to do some research. [30:51] It generated some plots on the right side panel, [30:54] and then finally it also generated [30:56] a cool artifact for me. [30:58] It even generated an infographic using Nano banana Pro. [31:02] It generated some graphs using the code it just wrote, [31:05] and then afterwards it generated some cool tables. [31:08] As you can see, Antigravity 2.0 is unabashedly agent first [31:12] and has been optimized to be the best surface for you to interact [31:16] with agents. [31:18] Let's take a look at the previous conversation [31:20] to see how it's going. [31:22] Antigravity ended up doing a whole host of research, [31:25] ended up writing over 100 lines of code, [31:27] and then finally built the operating system. [31:30] Let's take a peek and see if it works. [31:33] Moment of truth. [31:35] Amazing that never gets old. [31:48] Well, playing doom on an operating system [31:50] that Antigravity built is both fun and impressive. [31:52] It hasn't stopped there. [31:54] We've tasked these agents to build a photo editing [31:57] suite, a real time messaging app, [31:59] and a multi-user collaboration platform, [32:01] all with the same results. [32:03] Multi-day engineering efforts are collapsing [32:06] into hours, if not minutes. [32:09] This was made possible by the New subagent teamwork [32:11] capability, and we're excited to bring this [32:14] to you as an early preview in Antigravity. [32:24] Last but not least, 3.5 Flash is incredibly fast. [32:29] Sundar said, it's four times faster [32:31] than other frontier models. [32:33] But as we agentic coding is a token monster. [32:36] So we've taken it to another level in Antigravity. [32:39] We've optimized Flash to be not just four times, [32:42] but 12 times faster in Antigravity. [32:51] And we're thrilled to give you all a taste of this experience [32:54] starting today. [32:56] What we showed you today isn't just the vision, [32:59] it's how we're building Antigravity [33:00] to be the most complete agentic development [33:03] platform for everyone. [33:05] And we're doing it with the Google ecosystem. [33:08] Whether it's integrating tech stacks and tools [33:10] that you already use, or using Antigravity agent harness [33:13] to power the next crop of agentic [33:15] experiences across Google products. [33:18] Today, Antigravity 2.0 is available globally for everyone. [33:24] Join us in the developer keynote as we [33:25] demo all the new capabilities. [33:28] Back to you Sundar. [33:35] Thanks, Varun. [33:36] It's incredible that varun's entire OS [33:37] was built by a team of subagents in just 12 hours. [33:41] And for such a low cost, what's amazing about [33:44] Flash is how it delivers frontier level capabilities [33:47] at less than half the price of comparable frontier models. [33:51] We've heard that many companies are already [33:53] blowing through their annual token budgets, [33:55] and it's only May. [33:57] If companies used a mix of Flash and other frontier models, [34:00] they could save a lot of money. [34:02] To put this in perspective, top companies in Google Cloud [34:06] are processing about 1 trillion tokens a day. [34:09] If they shifted 80% of their workloads [34:12] from other frontier models to 3.5 Flash, [34:16] they'd save over $1 billion annually. [34:19] That is real savings they can pour back into their company. [34:22] Gemini 3.5 Flash is available for everyone today [34:27] across our products and APIs. [34:35] We are also excited for 3.5 Pro. [34:38] We're using it internally. [34:40] It's showing great improvements. [34:41] I know you can't wait to get your hands on it. [34:44] Give us until next month to get it to you. [34:47] Gemini 3.5 and an Antigravity are [34:50] unlocking a new world of agents and agentic capabilities. [34:55] We've been bringing agents to developers and enterprises [34:58] for a while now. [35:00] We are super focused on bringing the power of agents [35:03] safely and securely to consumers, [35:06] so that they work for everyone. [35:08] You will see agentic experiences across many of our products [35:12] today. [35:13] I'm particularly excited for what [35:15] we are bringing right into the Gemini app [35:18] introducing Gemini Spark. [35:21] It's your. [35:27] It's your personal AI agent that helps [35:30] you navigate your digital life. [35:32] Taking action on your behalf and under your direction. [35:37] It runs on dedicated virtual machines on Google Cloud [35:41] and it's 24/7 seven. [35:43] And yes, you can close your laptop. [35:49] It's powered by Gemini 3.5 and the Google Antigravity [35:54] harness, which allows it to perform long running [35:57] tasks easily in the background. [36:00] Spark will integrate seamlessly with tools, starting [36:03] with our own and in the coming weeks, with third party tools [36:07] through MCP, and you can work with Spark. [36:10] However, it's most convenient in the Gemini app [36:14] or soon through email and chat. [36:16] Let's have Josh come up and share more. [36:26] JOSH WOODWARD: Thanks, Sundar. [36:27] It's great to see everyone and let me show you [36:29] how Spark works with some examples from my personal life. [36:33] So here we are. [36:34] We have the new Gemini open. [36:36] It's been completely redesigned. [36:37] We'll talk about that later in the show. [36:39] And I want to take you into Spark here. [36:42] And what you can see immediately is a dashboard [36:45] with all the different tasks that I have [36:47] going on in the background. [36:48] It allows you to check in on these things. [36:50] And I'll paste in a task right off the bat. [36:53] This is a pretty straightforward example, but it's so useful. [36:57] Help me draft an email to the team. [36:58] Compile everything about our recent Gemini Live [37:01] launches and wins from the last week. [37:04] Use ghostwriter. [37:06] So there's a few things happening here. [37:08] Compile everything this goes across docs. [37:11] Email your chats and grabs the most important information [37:15] that you need for this update. [37:17] It also is going to use all of the stuff you [37:21] see in the time period of the last week with this ghostwriter. [37:25] That's a personal skill that I've written. [37:27] So the email sounds like me. [37:29] And what's great is that with Spark, you [37:31] can upload your favorite skills you find online. [37:34] So we're going to let this run in the background. [37:36] You can see it's already started doing various tool calls. [37:40] And I'm going to switch over to another one [37:42] from our personal life. [37:43] We're planning a big block party and you can see here, [37:46] this is a pretty complex prompt, we want [37:48] help grabbing all the RSVPs. [37:50] Keep a list of who's bringing what. [37:53] Remember to email those neighbors who [37:54] haven't signed up yet. [37:56] And what's amazing here is Spark will go through step by step, [38:01] look at all these steps all the time. [38:03] It saves you going through. [38:05] And again work across the various skills and apps [38:08] that you have. [38:09] And what's really amazing is it'll break it down and also be [38:12] able to generate files for you. [38:14] So the first one here, this is a live RSVP tracker [38:18] right in Google Sheets. [38:19] You can see that it shows who's confirmed and who hasn't. [38:22] What's amazing about this is it'll actually [38:25] update because it's connected to Gmail. [38:27] So when L. Thompson wrote eight RSVPs, [38:30] it'll update, which is pretty amazing. [38:33] The other thing is it keeps track [38:35] of all the different guests, as well as [38:37] sending follow up reminders to who hasn't signed up yet. [38:41] So again, this will happen. [38:43] It'll create the drafts and under my control, [38:45] I can send them. [38:47] And then finally, what the prompt did [38:48] is it created a hype deck for the block party. [38:51] And you can see creates right in Google Slides, [38:54] perfectly integrated. [38:55] And you can see it even pulls in things like our giant bounce [38:58] house that's going to be in the cul-de-sac. [39:00] Now, all of this is happening in the background [39:03] and under my control. [39:05] And what's amazing is Gemini can even go a step further and pull [39:08] out things like your neighborhood homeowners [39:11] association won't let you set this up before Friday [39:14] afternoon at June five. [39:16] That's pulling from a file in my Google Drive, [39:18] so it's incredibly helpful about how it pulls it all together. [39:22] Now that Spark working on a laptop. [39:25] Spark is also amazing on the Go, and it works [39:29] across both Android and iPhone. [39:31] So here you can see on my phone here I'll open it up [39:34] and I can go into Spark. [39:35] You see both of the tasks we just [39:37] went through they sync across all your devices, [39:39] which is so helpful. [39:41] And Spark is amazing at just brain dumping things [39:44] on your mind. [39:45] If you're super busy, it's almost you can just [39:47] throw tasks over your shoulder. [39:50] Spark will catch them and then run with them. [39:52] So watch this. [39:55] Start a few threads for me. [39:57] The first one. [39:57] Find all the upcoming meetings with Sundar [39:59] and turn them all hot pink so I don't miss them. [40:03] The second one. [40:04] Last night I met our new neighbor John. [40:06] Write a note to him and his family. [40:07] Invite them to our block party because they weren't on our list [40:10] originally. [40:11] The third one create a document with the top things my wife [40:14] and I need to do for the kids before the end of the school [40:16] year. [40:17] Categorize it by deadline and priority. [40:20] Make it easy to digest. [40:21] I don't want to miss anything. [40:24] All right. [40:25] So we'll send that in. [40:26] And you can see at the speed of my voice, [40:28] it's taking that one task. [40:30] And it will capture all of that context as fast as I can talk. [40:35] It starts out as a single thread here and in the background, [40:37] it's is actually going to go through and break those down [40:40] into individual tasks. [40:42] Now I can just put my phone away and get on with my day. [40:46] And Spark works in the background for me. [40:48] We'll check in later, see how it's doing. [40:57] You can see this is one of the first times we've [40:59] been able to put a phone down and let it [41:00] keep working on the I/O stage. [41:02] It's great because we're prioritizing safety. [41:05] We're rolling out Gemini Spark deliberately [41:08] to trusted testers this week. [41:10] And as a beta 4 us Google AI Ultra subscribers next week, [41:15] we want this new type help to be in as many hands as possible. [41:20] So to do that, we're introducing a new Ultra plan for $100 [41:25] a month. [41:32] And for those of you that need maximum limits, [41:35] we're dropping the price for our top tier Ultra plan [41:38] from $250 a month to $200 a month. [41:45] And there's so much more to come later this summer. [41:48] Gemini Spark will operate directly within Chrome, [41:51] acting as your agentic browser across the web. [41:55] It can take action on your tasks under your direction. [41:59] We're also building a dedicated home base for your agents [42:02] on your phone, Android halo, which is coming later this year. [42:07] As Sundar said, we've entered a new agentic era across Google [42:12] and we can't wait to see what you're going to build with it. [42:14] Back to you, Sundar. [42:21] >> SUNDAR PICHAI: Thanks, Josh. [42:22] So great to see Gemini Spark getting [42:24] things done on your behalf. [42:26] I've played around with all sorts of agents [42:28] and you can really see the potential, [42:31] but it's still early days when it [42:32] comes to making agents easy to use, [42:35] super secure and truly helpful. [42:37] That's why I'm really excited by Gemini Spark. [42:41] We are laying the foundation to bring this [42:43] all together in a safe and secure way to consumers [42:45] everywhere. [42:46] Look forward to having you all try it. [42:49] We are firmly in our agentic Gemini era. [42:52] Gemini Spark is the first experience [42:55] you're seeing made possible by 3.5 models and Antigravity. [43:00] This combination gives us new ways to accelerate our mission [43:04] and to transform our products to be radically more helpful, [43:08] and I can't wait to see how it will transform. [43:11] Search our ultimate moonshot this past year [43:15] has proven just how much innovation [43:18] is possible at the heart of our information mission. [43:21] And as we enter this agentic era, [43:24] search will be more helpful and powerful than ever before. [43:28] Let me turn it over to Liz to share what's next. [43:42] LIZ REID: People bring billions of questions to search every day. [43:47] Now, sometimes the whole world is searching for the same thing, [43:51] but more often your questions are just [43:53] as unique as you are to make it possible to ask whatever [43:58] is really on your mind. [44:00] Now to unlock this. [44:02] We've been on a journey to bring together the best of a search [44:04] engine with the best of AI, and we started this transformation [44:09] with AI Overviews. [44:11] And if you can believe it, it was just last year on this stage [44:14] that we launched AI Mode. [44:16] Now it's our most powerful AI Search, [44:19] bringing in our most advanced Gemini models. [44:21] And as of today, we're upgrading it on Gemini 3.5. [44:31] Now, as Sundar mentioned, AI Mode [44:33] has surpassed more than one billion monthly users, [44:36] and we're seeing phenomenal growth [44:39] with AI Mode queries more than doubling every quarter [44:41] since launch, as people have learned [44:44] how much more search can do. [44:46] They've started bringing more questions. [44:49] So much so that last quarter, we saw search queries [44:52] reach an all time high. [44:53] But what's even more remarkable. [44:56] You're asking your real questions. [44:58] In all their Super specific and detailed glory. [45:01] Knowing search can really tackle them. [45:04] You're having real conversations with search going back and forth [45:08] and deeper. [45:10] So you're not just asking nearby hikes. [45:12] You're asking, can you build an itinerary for hiking day [45:15] trip near me with great views and dog friendly trails [45:19] and a lunch spot with convenient parking. [45:22] And now we're entering the next chapter of Google Search. [45:25] We're incredible. [45:26] AI features aren't just in search. [45:28] Google Search is I search through and through. [45:32] Now it's an AI Search that brings together [45:34] our most advanced Gemini models, our newest agentic capabilities, [45:39] and the full breadth of the world's information. [45:42] With over a billion facts updated [45:44] a minute billions of new web pages indexed every day, [45:48] and connections to an infinite range of human perspectives. [45:53] So whatever's on your mind, you can come to Google [45:55] and truly ask anything. [45:59] Now, to start, I'm excited to announce [46:01] we're launching a brand new intelligent search box. [46:05] Before, the search box was a contained space, [46:08] but now it's totally reimagined. [46:09] With AI. [46:11] It expands with your curiosity and as you ask, [46:15] search helps you formulate your questions [46:17] with AI powered suggestions. [46:20] This goes beyond autocomplete. [46:22] It offers nuances that you might not have even thought to add, [46:26] helping you take the exact question on your mind [46:28] and ask it with ease. [46:31] This new search box puts our most powerful AI tools [46:33] right at your fingertips, and you [46:36] can ask across modalities with text, images, files, videos, [46:40] and search reasons across them all. [46:43] Now, this is the biggest upgrade to our iconic search [46:45] box since its debut over 25 years ago, [46:48] and it's starting to roll out today. [46:58] Next, we're making it even easier [46:59] to continue the conversation with search, [47:02] bringing AI Overviews and AI Mode into one seamless [47:05] AI Search experience so you can flow effortlessly [47:08] from your question to your response on the main search [47:11] results page to follow-ups in AI Mode, [47:14] and your context stays with you. [47:16] And as your conversation gets deeper, [47:17] your links and your sources get even more relevant [47:20] to what you might want to explore. [47:22] So you continue to get the best of AI and the best of the web. [47:27] And I'm excited to share this new seamless [47:29] AI Search experiences live today across desktop and mobile [47:33] worldwide. [47:41] Next, you just heard Sundar and Josh share our approach [47:44] to agents and the potential that this can open up. [47:47] Now we're taking an exciting step [47:48] toward this vision where you'll be able to create and manage [47:51] multiple AI agents for your many tasks in search. [47:56] We're entering the era of search agents now. [47:59] To start can set information agents to work for you 24/7 [48:03] in the background. [48:05] They can find you exactly what you need, [48:07] exactly when you need it, and help you take action. [48:11] You can spin up multiple agents and search simultaneously [48:14] to get updated and make progress on all those things [48:17] that matter to you. [48:19] And these will work with and alongside Gemini Spark [48:21] to help you get more done. [48:23] So let's put a few to work. [48:25] Say you're really into finance and you [48:27] want to know more about big biotech [48:28] want to know with P/E under 15. [48:31] Positive cash flows and low debt when it matters. [48:35] Now you can just ask all that and your agent is off. [48:39] It'll take your super complex question and map out a plan. [48:43] It determines the urgency. [48:45] Getting that you really need in the moment. [48:47] Intel and it sits triggers to look out for information as it [48:50] changes and picks the tools and data hooks it needs for the job. [48:55] It connects directly to our real time finance data, [48:58] so you get up to the second updates on stock prices [49:01] and insights on the market the moment it moves. [49:05] Now, when it does, your agent sends you [49:07] an intelligent and synthesized update. [49:09] It helps you understand what's going on so you can separate [49:12] the signal from the noise. [49:15] And it points you to Hyper relevant content [49:17] like this crowdsourced research platform, news sites and social. [49:22] This helps websites and creators get fresh content discovered [49:25] by people who really care about it, when [49:28] it matters most to them. [49:31] Now, let's say your apartment hunting. [49:33] You can do a total brain dump of what [49:34] you're looking for with all your criteria [49:37] like location and natural light and availability, [49:40] and your agent will continuously scan the entire web [49:43] across sites, social and forums. [49:46] Or if you're a sneaker fan, you can just [49:48] ask to be updated when any of your favorite athletes [49:50] announce sneaker collabs or drops. [49:53] And it monitors everything from blogs to our shopping graph [49:56] so you don't miss out. [49:59] Now you'll be able to put information agents [50:01] to work for you this summer. [50:03] Just ask search to keep you updated on whatever [50:05] you want to know. [50:08] Now, information agents are among the first [50:10] of many agents we're introducing in search to make [50:13] it more helpful for you. [50:15] So whether you want to find it, check it, book it, buy [50:18] by it or more search will be able to get it done. [50:23] Now we're also bringing agentic coding to search. [50:26] So it can build you custom experiences just [50:28] for your questions. [50:30] To show you how this works, here's Robbie. [50:44] ROBBY STEIN: We believe the best version of search [50:47] is one created just for you. [50:50] It's a search that gives you information [50:52] in the most helpful format for your question. [50:55] And we've spent years perfecting this. [50:58] So if you're shopping, we give you products asking about data. [51:03] You see charts looking for inspiration. [51:07] You get beautiful visuals. [51:10] Now we're taking this to a whole new level. [51:14] We're bringing Antigravity and the agentic coding capabilities [51:18] of Gemini 3.5 Flash right into search, [51:22] so search can build you the ideal format exactly [51:26] for your question. [51:27] Completely custom on the fly. [51:31] We're talking dynamic layouts, interactive widgets, [51:35] entire experiences all created just for you. [51:39] This is agentic coding at the scale of search. [51:50] Let me give you an example. [51:52] Say I'm a college student trying to wrap [51:54] my mind around astrophysics. [51:56] I can go to search now and just ask, how do black holes affect [51:59] space time. [52:00] And check this out. [52:01] I get this interactive visual right in the eye. [52:03] Overview search now gets for a concept this complex. [52:07] I need to interact with it to really understand it. [52:11] But this is still kind of 101. [52:13] So I'm going to follow up and say, now show me [52:16] how two orbiting objects like binary black holes [52:19] create gravitational waves. [52:22] Search dynamically builds a brand new interactive visual [52:26] in real time. [52:27] Completely custom for my specific question. [52:36] And now I can play with the parameters [52:38] like orbital separation and mass ratio. [52:41] This is so cool. [52:42] I can see how wave patterns have changed. [52:44] And now the smaller black hole spirals around the bigger one. [52:49] And now that I get the basics, I can dive into these resources [52:52] here and read things like the LIGO discovery papers [52:55] to learn more. [52:57] Now, you might be wondering, how can search actually [53:00] build custom UI like this for billions of unique questions. [53:05] With Gemini 3.5 Flash search plans the ideal response [53:10] from scratch. [53:12] It designs the layout, decides what custom components [53:15] to build fans out to do the research, [53:17] and then finally deploys the code [53:19] to build custom components like these in the response. [53:22] Search invokes in a gentle coding harness [53:25] powered by Antigravity so it can read and write files and execute [53:30] code in a secure containerized environment. [53:33] This is the same tech you saw Varun build a whole OS with, [53:37] and we're bringing that power right in to search generative [53:42] UI with Antigravity is rolling out to search this summer [53:45] for everyone free of charge. [53:54] So whatever it is you want to understand, [53:56] whether you're wondering how your watch actually [53:58] works, analyzing the cost of a new commute, or so much more, [54:02] you'll get responses as unique as your questions. [54:07] Now let's take this a step further. [54:09] Some projects aren't one off questions, [54:12] they're ongoing tasks. [54:13] So search now has the ability to help you create entire custom [54:18] stateful experiences tools, trackers, dashboards. [54:23] I think of these building my own little mini apps with search. [54:27] And I've been making a ton of them, [54:29] and they're especially awesome for those long running tasks [54:32] where you want to keep coming back, like planning a wedding [54:35] or managing your home move. [54:37] So what do you think. [54:39] Should we build one together. [54:40] Yeah All right, let's do a demo. [54:45] So I'm constantly trying to figure out [54:47] on Thursday what it is I should do with my family every weekend. [54:51] And so here's a search I actually [54:53] just did about fun things to do with my family. [54:55] This weekend. [54:56] And of course, you get great, great response here [54:58] from AI Mode. [54:59] But here's something new. [55:01] Search proactively offers to build a weekend planner for me. [55:04] And similar to how you just saw search, [55:06] create these generative UI and interactive visuals [55:09] completely from scratch. [55:10] Search can code this up right now. [55:13] Let's do it. [55:15] All right, so for the sake of today's demo, [55:18] we thought you might want to see a peek under the hood. [55:20] So as this builds you're going to see real time thinking steps [55:23] and code generation flowing in. [55:25] And here search is thinking about the right components [55:28] to build, not just what information to fetch [55:30] but the best way to present it. [55:32] And I've chosen to securely connect [55:34] Gmail, photos, and calendar. [55:36] So it's using personal intelligence [55:38] to make my suggestions even more helpful. [55:40] That means referencing Gmail, receipts, calendar, [55:43] and much more to personalize. [55:44] And boom, here it is. [55:45] Looks like it's ready. [55:48] All right, got a beautiful planner [55:50] already taken into account. [55:51] Driving times and weather and search [55:54] knows that I have two kids and that they love animals [55:57] and that my oldest is learning to play chess. [56:00] So actually the second one here is awesome for my oldest. [56:03] So I'm going to go ahead and heart that. [56:05] But two kids got to make them both happy. [56:09] So I'm going to lock in the Happy Hollow park and zoo right [56:13] there. [56:13] And because this syncs with my calendar, [56:15] you can see it's already blocked off. [56:17] My afternoon to meet a friend to watch this game down here. [56:21] And below it's got all of these cool restaurant reservations [56:23] and beautifully laid out on Maps. [56:26] Now, having seen this a little bit, [56:27] I kind of want the map to be higher up. [56:30] And also, my wife and I try to do Friday date nights. [56:33] So I'm going to keep customizing this. [56:35] And just for the sake of speed I'm going to paste in here one [56:39] more question to add Friday night date each week [56:42] and move it to the top. [56:44] And just like before, it's thinking through what's [56:46] needed to adjust the planner, looking up [56:48] real time information, even checking my preferences again. [56:51] Wow, that was so fast. [56:52] So it's able to use all kinds of information [56:54] from Google to build this on the fly. [56:57] And now you can see the map right at the top and Friday date [57:01] night tab right here. [57:04] So I can scroll down. [57:05] I can see all of these really great places [57:07] after the babysitter arrives. [57:08] Here's some awesome restaurants. [57:10] I'll go ahead and select it and we're ready to go. [57:12] Now, a weekend plan is not complete [57:15] until I get my wife Danielle's seal of approval [57:18] so I can share this app with her, which I'm going [57:21] to do right here and copy it. [57:25] And now I'm actually going to pull it up on a phone. [57:27] So when I send this to her, this is exactly what she's [57:29] going to see on her phone. [57:37] All right. [57:38] So Danielle is in, but it also looks [57:41] like she might have some feedback from me when [57:43] I get home. [57:43] But we're going to deal with that kind of a little bit later. [57:46] So all I have to do is say, add this to my calendar search, [57:50] will add it to all of our calendars, [57:52] and we're ready to go. [57:53] And next weekend I'll be able to come back and plan [57:55] a whole new weekend with my family just like this. [57:58] So we're bringing Antigravity into search [58:00] so you'll get generative UI this summer. [58:03] Starting with subscribers, you'll [58:05] be able to custom build experiences just [58:07] like this in the coming months. [58:17] From search agents to agentic coding. [58:19] This is an AI Search that does more for you. [58:22] Whatever your question, agentic capabilities [58:25] are going to transform all the ways you use search, including [58:28] how you shop to tell you more. [58:30] Here's video. [58:45] VIDHYA SRINIVASAN: With the incredible advancements in AI. [58:48] We are entering a whole new era. [58:51] We've been building the foundation for agent e-commerce [58:54] and now we are bringing that future right to you. [58:58] People shop across Google over a billion times a day. [59:03] And for years, we've connected you with brands and retailers [59:07] to help you get exactly what you're looking for. [59:10] It starts with our shopping graph, [59:12] the world's most comprehensive catalog of products. [59:15] It now has over 60 billion listings, [59:18] and they're constantly updated. [59:21] Combine the scale of the shopping [59:22] graph with our advanced Gemini app models, [59:25] and you get entirely new ways to shop. [59:28] It's more powerful, more intelligent. [59:30] Add more fun. [59:32] Now, when it comes to agent commerce, [59:35] we are focused on delivering three key building blocks [59:39] to make this future a reality. [59:41] First, the universal commerce protocol, or UCP for short. [59:47] UCP does for agent commerce what HTTP did for the web. [59:53] It gives agents and systems a common language. [59:57] It's an open source standard that [59:59] allows all the key players to work across the entire shopping [60:03] journey. [60:04] That means it makes everything from product research [60:08] to checkout to shipment tracking totally seamless. [60:12] And we've been thrilled to see the entire industry [60:15] rally behind it. [60:16] We co-developed and launched UCP with an incredible group [60:20] of founding partners and recently welcomed [60:24] Amazon, Meta, Microsoft, Salesforce, and Stripe. [60:36] They're going to help steer this as the open standard. [60:41] It may very well be the first time [60:42] we've all agreed on something. [60:47] And now we are bringing UCP to even more verticals like hotels [60:52] and local food delivery providers, and to YouTube [60:56] and more products. [60:58] We're also expanding UCP powered experiences on Google products [61:02] to more regions like Canada, Australia, [61:06] and the UK in the coming months. [61:10] That brings us to our second building block the agent [61:14] payments protocol, or AP2. [61:17] So when it comes to agent payments, [61:19] the number one question we hear is, [61:22] how do I know it just won't go off and buy something [61:25] I don't want. [61:27] Motorcycle anyone. [61:30] It's a fair question and it's why we created AP2. [61:34] It's designed so your AI agent can securely [61:37] make payments on your behalf, but always under your control. [61:42] For us, it comes down to two things [61:44] setting boundaries and ensuring accountability. [61:48] First, let's talk about boundaries. [61:51] We are now making it easy to set strict guardrails. [61:55] Just tell your agent the specific brands and products [61:58] that you want and how much you want to spend. [62:02] It automatically makes the purchase [62:04] if the criteria are met, but boundaries only [62:08] work if there's accountability. [62:10] Ap2 creates a transparent, verifiable link [62:14] between you, the merchant, and all the payment processors. [62:18] It uses privacy preserving technology [62:21] to keep your data safe and with tamper proof digital mandates. [62:25] Ap2 ensures the agent is always acting on your behalf [62:30] and gives you a permanent digital paper trail. [62:33] So if you ever need to make a return, you and the merchant [62:36] are looking at the same record. [62:39] Your payment info stays shielded, [62:42] the data stays private, and your purchases stay secure. [62:47] We'll begin bringing AP2 to Google products in the coming [62:50] months, starting with Gemini Spark. [63:00] So these protocols are powering the foundation of this new era. [63:04] Which brings us to our final building [63:06] block of agent e-commerce. [63:09] I am excited to announce the universal cart, a truly [63:13] intelligent shopping cart. [63:16] It works across merchants and across services. [63:20] You'll be able to add things to your cart when you're browsing, [63:23] search, chatting with Gemini app, [63:25] watching YouTube, or even reading your Gmail. [63:29] The moment you add a product, your card [63:31] goes to work for you in the background. [63:34] It finds deals, looks at price drops, [63:37] gives you insights on the price history, [63:39] and alerts you when something comes back in stock. [63:42] It all runs on our Gemini models, [63:45] so your card gets even smarter as the models improve. [63:50] So just think of it as shopping with superpowers. [63:55] Now another game changer is how it [63:56] applies intelligent reasoning. [63:59] Let's say you're building your first custom PC. [64:02] You see a motherboard with great reviews [64:04] and just add that to your cart. [64:06] Now you already picked out a processor, [64:08] but what you didn't realize is that the processor [64:11] needs a motherboard with a different type of socket. [64:14] Your cart catches this for you and suggests an alternative. [64:18] Preventing a problem you didn't really see coming. [64:22] Next, my. [64:24] My favorite feature is how the cart can actually [64:26] find hidden savings for you. [64:29] If you're like me, you have different payment cards [64:32] with different perks that are just hard to keep track of. [64:35] Now you don't have to. [64:37] The cart can do it for you because it's [64:39] built on Google Wallet. [64:41] Here's an example. [64:43] Here there's an offer from target for some products [64:47] that were added to the cart earlier in the week. [64:51] I see it, so I just want to get it now. [64:53] And UCP makes us checkout from your cart. [64:56] Super smooth for many of your favorite brands, [64:59] you can check out right on Google in just a few taps [65:03] with Google Pay. [65:04] Or you can transfer all these items straight to the retailer's [65:07] site and then buy it there. [65:10] I am excited to announce that we are rolling out [65:13] the universal cart in the US across search and the Gemini app [65:17] this summer, with YouTube and Gmail to follow. [65:29] The building blocks we showed you today won't just [65:32] create the foundation for e-commerce, [65:34] they will transform the way you shop on Google, making it more [65:39] powerful, more intelligent. [65:41] And a lot more fun. [65:43] Happy shopping and back to you, Liz. [65:51] Today, you've seen how we're bringing together [65:54] the best of a search engine with the best of AI [65:57] to build a Google Search that's more helpful [65:59] and powerful than ever before. [66:02] We're search agents. [66:03] Work for you around the clock. [66:06] Wear agentic. [66:06] Coding means that search can build experiences [66:09] as unique as your questions, and wher [66:12] the power of agentic commerce isn't in the distant future, [66:16] it's here now. [66:19] So in this chapter of AI Search, Google [66:22] can help you find, understand, build, do anything. [66:27] So go ahead, just ask. [66:47] But when I'm with you, baby, I go out of my head. [66:51] And I just can't get enough. [66:53] I just can't get enough. [66:55] All the things you do to me. [66:57] And everything you said. [66:58] I just can't get enough I just can't get enough. [67:10] We walk together. [67:12] We're walking down the street. [67:14] I just can't get enough I just can't get enough. [67:18] Every time I think of you, I know we have to meet. [67:21] I just can't get enough. [67:23] I just can't get enough. [67:33] And when it rains. [67:34] You're shining down for me. [67:36] And I just can't get enough I just can't get enough. [67:40] I just can't get enough I just can't get [67:43] enough I just can't get enough. [67:45] I just can't get enough. [68:01] JOSH WOODWARD: All right. [68:02] Hi, everyone. [68:02] It's great to see you again. [68:04] And you've already seen some amazing breakthroughs [68:06] this morning. [68:07] And we're just getting started. [68:09] We have a lot more to show you in the Gemini app, [68:12] so let's jump right in. [68:14] Over the last summer, Gemini app has had incredible momentum. [68:17] More than 900 million users are coming to the Gemini app [68:22] every month, and a big part of that growth [68:25] is how fast we're shipping a year ago, none of these features [68:29] existed, and now Gemini has become the Ultimate [68:33] creative tool for everyone. [68:36] You can create images, videos, and music in ways people [68:39] never thought were possible. [68:41] We've made Gemini more tailored with personal intelligence, [68:45] allowing you to securely connect your Gmail, photos, [68:49] and other apps so you get customized help. [68:52] And just last week, we expanded personal intelligence [68:55] to users around the world directly in the app. [68:59] Millions of people are using it every single day. [69:02] They found it so helpful for things like personalized product [69:06] recommendations or trip recommendations, [69:09] or acting as a thought partner for navigating [69:12] big decisions in life, like a career change or in my case, [69:16] finding the right set of tires for our super cool minivan [69:19] that's going to be at that block party. [69:21] We've made Gemini superb for students [69:24] also over the last year. [69:26] You can use it to create guided learning, [69:28] generate practice tests, and even use dedicated notebooks [69:32] to keep all your notes and assignments in one place. [69:36] And these same notebooks sync directly to NotebookLM, just [69:40] on its own. [69:41] Notebook Elm has now been used to create [69:44] more than 1.5 billion notebooks, podcasts, slide decks and more. [69:50] Turning complex information into easy to understand knowledge. [69:55] And as you heard from Sundar earlier, [69:57] we're rolling out our new Gemini 3.5 Flash [70:00] model, which is unlocking a new set of features [70:03] that I'm going to show you today. [70:05] All of this is happening in a Gemini app that's [70:08] now available in more than 230 countries and over 70 languages, [70:14] making Gemini the most widely available AI [70:17] assistant in the world. [70:24] Gemini is becoming that universal assistant people [70:27] are turning to. [70:28] It's designed for their everyday lives. [70:31] And speaking of design, that brings us [70:32] to the first of our three big updates today. [70:36] Today, I'm excited to announce that we've completely redesigned [70:40] the Gemini experience from the ground up, [70:43] from the moment you open it. [70:44] We've greeted you with a stunning new design language [70:47] we call Neural Expressive. [70:50] We've added fluid animations, vibrant colors, new topography, [70:55] haptic feedback throughout the app. [70:58] But we all know that good design isn't just about how it looks. [71:02] Good design is about how it works. [71:05] So we've evolved the entire experience. [71:08] We've made it easier to discover and generate [71:10] those gorgeous images, videos, and music [71:13] with built in templates that you can easily remix. [71:16] And we've completely transformed the Gemini Live experience. [71:20] It now opens up immediately and inline, [71:23] and soon you'll be able to pick a regional dialect that [71:27] resonates with you. [71:30] You've got a right good mix of different accents knocking [71:33] about this one from Liverpool. [71:38] [Speaking various languages] [Speaking various languages] All [71:52] right, pretty cool. [71:57] We'll be rolling out these and many more regional dialects [72:01] in the coming weeks. [72:02] And my favorite part is how we handle model responses [72:06] with our new Neural Expressive design language. [72:09] That's where it really comes to life. [72:11] You won't see a wall of text anymore. [72:14] Instead, Gemini will carefully lay out its response [72:17] in real time just for you. [72:20] The generative UI you saw in search, as you scroll, [72:24] you might see interactive images that [72:26] are custom generated by Gemini. [72:28] You can drill into them like this [72:30] and explore information on an entirely new level. [72:33] You might even see timelines that you can quickly [72:36] skim or embedded videos. [72:38] So whether you're in dark mode or light mode, [72:41] the entire experience feels fluid and futuristic [72:45] and incredibly natural. [72:48] Best of all, Neural Expressive is rolling out globally [72:52] on Android, iOS, and the web starting right now. [73:01] With a completely redesigned Gemini app as the new base, [73:04] we can go even further with Gemini's creative abilities. [73:08] And that's update number two. [73:10] Gemini Omni is coming right into the Gemini [73:13] app for paid subscribers today. [73:15] It's going to let you bring your ideas [73:17] to life using any combination of text, images, and video inputs. [73:22] And as I've been using it, it feels [73:24] like the Nano banana for video moment is here. [73:28] It's never been so easy to create and edit videos. [73:32] Let's look at how this plays out in the real world. [73:35] I want you to meet Sasha. [73:36] She's working on a new song and she wants [73:39] to create a quick video teaser. [73:41] So she shares the raw video. [73:44] She adds some reference visuals to it. [73:46] And now this is the coolest part. [73:48] She can transform the style of her video, [73:51] maybe even switch the camera angle to a 360 degree shot. [73:55] And Gemini puts it all together for her. [73:58] Let's take a look at what it looks like. [74:09] All done within Gemini app. [74:14] As you can see, Omni understands the physics of her movement. [74:18] It intelligently layers those effects into the real world [74:21] without losing the soul of her shot. [74:24] Creating, remixing, editing videos has never been easier. [74:28] Year so whatever vision is in her head, [74:31] she can now use Gemini to make it real. [74:34] Google AI plus Pro and Ultra subscribers around the world [74:39] can try Gemini Omni today right in the app. [74:48] The third update today is about how agents are coming to Gemini. [74:52] This is a big shift for Gemini because agents don't just [74:55] answer questions, they proactively work on your behalf. [75:00] And to show you what this means, I [75:01] want to introduce one of our newest out of the box [75:03] agents called the Daily brief. [75:06] It's a personalized digest that's [75:08] designed to be your first stop every morning. [75:11] Here's how it works. [75:13] You can see here that it's synthesizing information [75:15] from across my inbox, my calendar, my tasks. [75:19] It's finding the most important things for me to be aware of. [75:23] And I totally can forget these things [75:24] like that returning that book. [75:26] It's not just summarizing the data, though. [75:29] It's organizing it by topic. [75:31] It's even suggesting the next step. [75:34] And with this travel info, I can just take the next step right [75:37] in line. [75:39] All of it is super concise in this morning digest [75:43] that's built for skimming. [75:45] And I can't wait for you to try it out. [75:47] The Daily brief is rolling out today, also [75:50] to Google AI plus Pro and Ultra subscribers starting right here [75:54] in the US. [75:55] And that's just to help. [75:56] You're going to get before breakfast. [76:00] Beyond the daily brief, we're allowing power users [76:03] to create their own custom workflows with Gemini Spark, [76:06] like I showed you earlier. [76:08] You'll remember at the beginning of the show [76:09] we sent a few tasks off. [76:11] So let's go check in on them, see how they've completed. [76:15] And you'll remember when I did this earlier. [76:17] I did it from the phone. [76:18] So we're going to actually pull it up on the laptop [76:20] to show how it syncs across. [76:22] And if I open up your eyes and the laptop, you can see it [76:26] subdivided those three tasks right here 1, 2, 3. [76:30] And it's actually got a really nice feature. [76:32] There are certain tasks they'll actually ask for your input. [76:36] So you can approve it. [76:37] So it doesn't just go off and do things you don't want. [76:39] But I'll check in this one here. [76:41] This is our school year planning checklist [76:43] and I asked it remember to create [76:44] a doc of all the things I didn't want to forget between now [76:47] and the end of the year. [76:48] And so I'll open this one up. [76:50] And what's amazing about it takes advantage [76:52] of all the Google Docs formatting, so I can immediately [76:56] click in and see the checklist here for our various kids, [76:59] all subdivided like this. [77:01] And easily to go one by one. [77:04] The date, the activity, the color [77:05] coding all integrated in one spot. [77:09] Pretty incredible how much time this can save. [77:12] Now what comes across all of these cases [77:16] is that we want to make agents easy to use and safe and secure. [77:21] As a reminder, Spark is are going [77:22] to start rolling out to trusted testers this week, [77:24] and we're planning to roll it out [77:26] as a beta to Google AI Ultra subscribers next week in the US. [77:31] We'll be bringing a version of Spark [77:33] also to Gemini Workspace to Workspace [77:36] as well as Gemini enterprise. [77:38] And this is just the beginning. [77:40] We've got a packed roadmap of features [77:42] we'll be shipping throughout the summer. [77:45] It's going to be amazing. [77:46] I'm really excited about the integrations that's [77:49] going to enable Spark to handle so many tasks in ways [77:53] that are even more proactive and powerful. [77:56] Imagine Spark will be able to look ahead on your calendar [77:59] and see that you're on snack duty for your kid's tee ball [78:02] game on Saturday. [78:04] It can go ahead and proactively set up your Instacart order [78:07] on its own. [78:08] So you don't forget those snacks. [78:10] And it'll even remember to pick snacks [78:12] that don't have nuts in them. [78:14] And we have this incredible lineup [78:16] of partners that are coming that will be integrating [78:18] with Spark over the summer. [78:22] Now I know I've promised just three updates, [78:24] but we have one more. [78:26] Last month, we dropped the Gemini app for macOS. [78:30] Here it is on the screen. [78:31] It's gorgeous. [78:32] This was a small team that built this native app from scratch [78:36] using Antigravity. [78:37] They did over 100 features in less than 100 days. [78:42] Now, two weeks ago, we were hacking on the Mac App [78:44] and we came up with something cool [78:46] and we wanted to sneak it into the show. [78:48] You all want to see it live. [78:51] All right. [78:53] We've got a big summer trip coming up, [78:55] and we've got to find a kennel for our two dogs. [78:59] And here's a picture of our two dogs. [79:01] There's Hank looking good, and Louis cinnamon, one [79:05] of the most interesting names for a dog we've ever heard. [79:08] What we're going to do and remember [79:10] when you have to go to a new kennel, [79:11] there's lots of paperwork, allergies, vaccines, [79:14] all the history you have to pull together. [79:17] It's so painful. [79:18] And so what you can do with this on Gemini on Mac OS [79:23] is actually take a look at a bunch of documents like this. [79:26] You'll be able to select them all [79:28] and then long press the function key [79:31] and just dictate the email to the kennel. [79:34] So it works something like this. [79:38] Hi there. [79:39] I need to do a shortboarding stay [79:41] for my two dogs, Louis Cinnamon and Hank, [79:43] starting this Thursday. [79:45] Oh, wait, no, actually, it's this Friday. [79:47] They've never stayed with you before, [79:49] but they're very social dogs. [79:50] And also, can you turn these files into a table [79:54] with their details. [79:55] Allergies, recent vaccines, and make this email [79:59] sound friendly so we make a good first impression. [80:04] All right I'm going to release the function key. [80:06] You can see Gemini's thinking at the bottom here on this MacBook. [80:10] What it's done is because I've selected those files in Finder [80:14] using its multimodal understanding. [80:16] It can go through the PDF, it can [80:19] go through these images of the invoices, [80:22] and it's all controlled by my voice. [80:24] So it can actually take all that complex information [80:27] and look at that. [80:27] There it is. [80:29] It's got a table in line. [80:35] It's also so amazing because it correct. [80:38] Remember I said Thursday. [80:39] No, scratch that Friday. [80:40] And it picks up that and automatically [80:43] cleans up my input. [80:45] This is the power of what Gemini can do using your voice. [80:50] These new voice capabilities and Gemini Spark [80:53] will be coming to the Mac App this summer as well. [81:01] And so there it is Today has been a jam packed Gemini day. [81:06] We've completely redesigned the entire experience [81:09] with Neural Expressive. [81:11] We're shipping the brand new Gemini Omni model and the 3.5 [81:15] Flash model with 3.5 Pro coming soon, [81:18] and you can now put Gemini to work [81:21] and it'll keep working for you even while you sleep. [81:24] Thanks to New features like the daily brief and Gemini Spark. [81:29] All of this gets us closer to our vision [81:31] of a universal assistant that's personal, proactive, [81:36] and powerful in your daily life. [81:38] So whether you're a student, a busy parent, or a small business [81:42] owner, we look forward to what you're going [81:45] to be able to do with Gemini. [81:47] Thank you. [81:49] Hi, I'm Holly. [81:51] I moved from South Korea to the US in 2005. [81:57] Food has always been something we love to do in our family. [82:01] So I decided to open a restaurant. [82:04] But running a restaurant is so hard. [82:09] I discovered Gemini could help me with a lot of things. [82:12] From menus to marketing, budgeting and inventory. [82:17] How much sesame oil did you order last month. [82:20] Last month you ordered five gallons. [82:23] I even rebuilt our website with Antigravity and Stitch [82:27] to include the custom chatbot for our customers. [82:30] But one day we needed to find the dishwasher. [82:33] Very last minute and that was my moment. [82:36] I wanted to help small businesses and people find [82:40] each other faster. [82:41] So I got a small team together, and we [82:44] used Gemini models to build an inclusive hiring platform [82:48] called work onward. [82:49] And we did what other job sites are not doing. [82:52] We translated the job postings and then made [82:54] it possible to post jobs just by text. [82:58] It helps break down barriers for those who aren't tech savvy [83:02] and help people find their place. [83:08] Our little tool is now a platform used by over 13,000 [83:12] people in New York City. [83:15] It's about honoring the dignity of workers and small businesses, [83:20] and I think that's the power of AI. [83:39] SUZ CHAMBERS: As you've heard today, our models and products [83:42] are unlocking new breakthroughs. [83:45] But the real breakthrough isn't the technology. [83:48] It's what you do with it. [83:50] Whether you're a designer, an entrepreneur, or an artist, [83:54] our products help shrink the gap between the moment [83:57] you have an idea and the moment you create it. [84:01] At its best, technology is a Canvas for human creativity. [84:05] And today, I want to dive into three products [84:08] that help you bring your ideas to life. [84:12] Let's start with one that takes the power of Nano banana [84:15] and gives you even more creative control. [84:18] Introducing Google picks a new product in Google Workspace. [84:24] Picks is our image creation and editing tool [84:27] that helps you create just about anything, from party flyers [84:31] to infographics. [84:33] With the creative controls you want, watch how easy this is. [84:38] You start with a base image as your Canvas, [84:41] and what's really cool is that picks understands [84:44] what's in your creations and how the objects work together. [84:48] You can hover over an element and click to remove it, [84:52] or you can resize an object to fit the frame. [84:56] Once the layout is set, you can add or edit text [85:00] and translate all of it with just a few clicks. [85:08] Pretty cool. [85:11] Every output from our creative tools, including pics, [85:15] is watermarked with SynthID and pics is rolling out this summer. [85:26] But what if you want to go beyond images. [85:29] Maybe you want to design an app or a website. [85:32] Now you can build UI at the speed of thought. [85:37] Teams across Google use our design product [85:39] called Stitch to turn rough ideas into beautiful UI designs. [85:45] But it's not just us. [85:47] In the last year, the world used Stitch to generate over [85:50] 100 million UI screens. [85:53] And to continue that momentum today, [85:56] we're introducing new ways to design. [85:58] Let's look at how this works. [86:01] My friends Tyler and Jenny own a pizza company [86:04] and they have a ton of experience making pizza. [86:07] Had no experience designing websites with a single prompt. [86:11] Stitch will go to work generating that UI. [86:14] Live now this is just the first pass, [86:18] but if they want to refine it, they can collaborate with Stitch [86:21] in real time, either by writing prompts or using their voices. [86:27] For example, they can jump in and say, [86:30] make the header text larger and update the menu [86:33] to highlight more pizza options and the layout updates [86:37] in real time. [86:39] And because Stitch links to many tools, [86:41] they can export their design to code or launch their website [86:45] in just a few clicks. [86:48] These updates in Stitch are rolling out today to users [86:51] globally. [86:58] With every new technology. [87:00] What's most exciting is seeing what people make with it. [87:04] That's why from day one, we haven't just [87:07] built models and tools for creatives, [87:09] we've built with them. [87:11] Let's take a look at some of our amazing partners [87:14] and what we've created together. [87:19] You're in the era where the human has [87:22] to be the most creative. [87:36] Toto, have a feeling we're not in Kansas anymore. [87:42] Yeah, we'll do that. [88:06] That spirit of collaboration is why we launched Google flow [88:09] at I/O last year. [88:11] Today, millions of people are using [88:14] it to create images, films and music [88:17] in ways they never could before and to build on that progress. [88:22] We're rolling out Gemini Omni, a new agent, custom tools [88:27] and music remixing. [88:37] Let's start with Gemini Omni. [88:39] Take a look at this raw footage. [88:42] I love how this person is walking, [88:44] his presence, his pacing. [88:47] Let's not change any of that with a simple prompt and style [88:51] reference. [88:52] Omni allows us to transform the environment, [88:55] add visual effects and any other element, [88:58] all while preserving the original performance. [89:02] And now you can even add new characters [89:05] while maintaining everything else in the scene. [89:15] Next, let's look at our second big update. [89:17] A new agent in Google flow. [89:20] Until today, flow could only execute one prompt at a time. [89:25] Now your agent can take multiple actions all at once. [89:30] Starting with just a single image, [89:33] I can ask the agent to help me find the best camera [89:36] angles for this scene. [89:38] It analyzes what's happening in the image concepts, the most [89:41] compelling angles, and then boom, [89:46] a single image becomes 16 unique videos. [89:55] The agent can also handle large scale edits, [89:58] like transforming all of the scenes from early morning [90:02] to late at night. [90:04] Its understanding of context is precise. [90:07] The desert sky goes completely dark [90:09] and the headlights turn on, illuminating the dust. [90:13] It's a true collaborator, helping [90:15] you create and edit at scale. [90:25] Our next update is flow tools. [90:27] Now you can vibe code any creative tool [90:30] you can think of right in flow. [90:33] Custom built by you for your unique creative process [90:37] like designing video effects, hand-drawn animations, [90:42] or layering text. [90:44] You can start building, sharing, and remixing tools today. [90:54] Visual magic is only half the story. [90:58] Google flow music brings the same creative control [91:01] to help artists create original songs for months. [91:05] One of our teammates had a piano riff in his head. [91:08] Let's listen to his original recording. [91:19] It's a cool foundation, but it wanted [91:22] to turn it into a demo to guide his band, [91:25] so he recorded his piano into flow music [91:28] and prompted it for an R&B direction [91:31] with a female vocal to inspire his band. [91:33] Singer let's take a listen. [91:46] If I'm not on display, tell me why would I try to fly. [91:51] Why would I try. [91:59] Now, this isn't his final track but it [92:02] helped his band decide what to record next. [92:05] And all of these features in Google, [92:08] all these new features in Google flow and Google flow music [92:12] are available today. [92:19] From musicians to small businesses and vibe [92:22] coders to artists. [92:24] The real breakthrough isn't the technology, [92:27] it's what you do with it. [92:29] And we can't wait to see what you create. [92:37] Next, I'd like to hand things over [92:39] to shiram to show you what happens when [92:41] we take Google's latest innovations [92:43] and bring them into the real world. [93:00] SHAHRAM IZADI: This is such an exciting time for XR. [93:03] AI continues to unlock all new experiences [93:06] on headsets, glasses, and everything in between. [93:11] Android XR, the new platform we built with Samsung [93:15] and optimized for Snapdragon with Qualcomm, [93:19] combines this pioneering hardware with Gemini. [93:23] This gives you help in the moment [93:25] without taking you out of it. [93:28] The next big milestone for Android XR [93:31] is intelligent eyewear. [93:34] There'll be two types of these AI glasses [93:37] that connect to your phone and give you hands free help all day [93:41] long. [93:43] Last year, we showed you display glasses on the I/O stage [93:48] with the small in-lens display, you'll get helpful information [93:53] right in front of you right when you need it, [93:56] seeing your Uber pickup details at a glance, [94:00] or getting live translations as you travel. [94:03] You'll even be able to use features like create my widget [94:08] to make glanceable elements. [94:11] The first wave of developers are already [94:13] creating display experiences, and you'll [94:16] hear more about these glasses later this year [94:19] when we expand our trusted tester program. [94:23] But let's talk about what's launching this year. [94:27] Today, I'm excited to announce that our first audio glasses [94:32] will arrive this fall. [94:40] They are designed to give you all day help with Gemini that [94:43] is spoken into your ear privately, rather than shown [94:47] on a display, and these glasses let you stay hands free [94:51] and heads up for things listening to music, [94:55] taking photos, making calls, or tapping into your phone apps [95:00] all without reaching for your pocket. [95:03] Personally, I love to cook, but I'm not one [95:06] to follow a recipe book, so it's great to have [95:09] Gemini offer some advice before I get too experimental. [95:14] These audio glasses have brought together [95:16] an all-star cast of partners. [95:19] You've got the world's top eyewear designers [95:22] at Gentle Monster and Warby Parker creating iconic designs. [95:30] Thank you. [95:34] The world's leading electronics company, Samsung, [95:37] is building innovative new devices and experiences that [95:41] set the bar for the whole industry, [95:45] and we've been working to bring the best of Google [95:47] to these glasses as well. [95:49] And yes, they're going to pair with both Android and iOS [95:53] devices. [96:00] They look incredible. [96:01] And today you are actually going to finally see for yourself. [96:06] Let's pass it to our friend and partner from Samsung, Jay Kim, [96:11] to kick off the world's first look. [96:16] At Samsung. [96:17] Our vision is to enrich people's lives [96:20] and help shape how we live tomorrow [96:23] in close partnership with Google. [96:25] We're introducing intelligent eyewear [96:27] that empowers you to connect to the world with confidence. [96:30] Built with Samsung's precise engineering and craftsmanship, [96:34] we're merging form, function and intelligence [96:38] to create something you will want to wear in eyewear. [96:41] Every millimeter counts. [96:43] Accounts today, we're thrilled to share a first look [96:46] at the upcoming styles co-created [96:48] with our eyewear partners, Warby Parker and Gentle Monster. [96:52] Let's take a look. [96:58] People always think that hey, we have to be disruptive, [97:02] but we have to think about what is [97:03] the disruption I want to make. [97:06] Intelligent eyewear that looks prettier than normal eyewear. [97:09] That's our goal. [97:11] It's the balance between technology and vision. [97:14] It's not only product, it's all about perception and emotion. [97:18] Hurt That is the DNA. [97:21] I want to give confidence to people. [97:23] When I wear this, they feel connected [97:26] with this kind of braveness, this kind of rebellious things [97:29] because this is who we are. [97:31] That's the meaning of Gentle Monster. [97:33] With Google and Samsung. [97:37] Glasses are a deeply personal product. [97:40] They shape how you see the world and how the world sees you. [97:43] What you see here is an evolution [97:46] of Parker's first intelligent eyewear designs. [97:50] Our vision for intelligent eyewear [97:52] started by wanting to design a beautiful pair of glasses. [97:56] We often take inspiration from different art objects, [98:00] artists eras in time. [98:03] It was important for us not to hide the technology, [98:06] but to celebrate it. [98:08] These will not only help people see, [98:09] but will help them understand the world more deeply. [98:12] We want people to experience the world fully [98:15] and that's the magic of this technology. [98:18] Hey, Gemini app, who wore it better. [98:30] Aren't they beautiful. [98:32] These are the. [98:33] These are the first two designs of a bigger collection that's [98:37] coming this fall. [98:39] What time it is now. [98:41] Who's up for a live demo. [98:46] Nishtha is going to join me on stage. [98:49] So let's have Gemini play some entrance music for her. [98:53] And it's seeing the world around [98:55] around you. [98:56] You can see. [98:57] So the more hyped you are in the audience, [98:59] the more pumping the music will be. [99:07] Hey Gemini! [99:10] I think it's done it already. [99:11] Play that music. [99:12] Come on, let's go. [99:17] Nishtha, welcome! [99:20] Welcome, Nishtha. NISHTHA BHATIA: Thank you. [99:23] I'm wearing one of the Warby Parker styles [99:26] and Nishtha has a pair of the Gentle Monster glasses [99:28] on looking great nishtar. [99:30] Thanks, Jerome. [99:31] You too. [99:32] OK a quick preamble for the demo. [99:34] Today we're going to be piping in the audio to the stage [99:39] speakers from our glasses, but in regular use, of course, [99:42] Gemini is just going to privately be talking to us. [99:45] Now, last time we were on stage, we [99:48] saw the rich Google Maps experience on display glasses. [99:52] Well, Maps on these glasses are just as helpful, [99:55] especially when working in combination [99:57] with personal intelligence. [99:59] No pulling out my phone. [100:01] All I have to do now is ask, hey, [100:06] can you navigate me to the place I met my friend Gianna [100:09] at last week. [100:11] Hey, I've set your route to the Redwood Grove nature [100:14] preserve from last week's hike. [100:16] Want to add a stop on the way to grab your afternoon cold brew. [100:19] Yes Gemini, I would love that. [100:22] OK, I'm starting walking navigation [100:24] with a stop at Coupa cafe. [100:26] Turn around and head towards Bill Graham Parkway. [100:29] With glasses, you can allow maps to further understand [100:33] your context and what's in front of you. [100:36] So you get detailed directions like the coffee shop [100:39] is coming up on your right. [100:40] Speaking of coffee, do you want to show [100:43] how Gemini intelligence can allow you to order [100:45] that cold brew up ahead. [100:47] That's a really good idea. [100:51] Gemini can you actually put my usual order [100:54] in at that coffee shop we just talked about. [100:57] Sure I'll order you a Nitro Cold brew for pickup [101:00] from Coupa cafe on DoorDash. [101:03] So on the screen behind me, you can actually see Nisha's phone [101:07] that's in her pocket. [101:08] Gemini is able to launch apps like DoorDash, [101:11] then click through all the different options [101:14] screens automatically to order her coffee. [101:18] And in a moment, it's going to be ready for confirmation. [101:26] There we go. [101:32] There we go. [101:37] I have prepared your order for the Nitro Cold brew [101:40] from Coupa cafe. [101:41] Would you like to confirm. [101:43] Yes, please and add a tip for 20% too. [101:47] I think after all these rehearsals. [101:49] That's our 15th cold brew today. [101:51] And thanks to the coffee shop for putting up with us. [101:56] Definitely is. [101:57] And speaking of preparation, I actually [101:59] muted my text before I came on stage, [102:01] but let me see if I can catch up now. [102:04] Hey, Gemini. [102:05] Any important messages I miss. [102:08] Yes your family group chat decided [102:11] to meet for dinner at o'clock today. [102:13] And there is a reminder to say your daily affirmations. [102:16] You are strong. [102:17] You are confident. [102:18] Affirmations a bit later, but why don't you [102:21] add that dinner to my calendar. [102:24] Sure, I'll add an event for family dinner at 7:00 PM [102:27] to your calendar right after your team celebration. [102:31] Awesome Gemini. [102:33] Not only. [102:34] Thank you. [102:37] Gemini not only summarized her muted text, [102:40] but it tapped into her calendar app to add that event. [102:44] So we saw how glasses work on your phone. [102:47] But what about when they're attached to your watch to give [102:51] you a glanceable display. [102:53] It's not a Google I/O without an audience selfie, [102:56] but we're going to do it with a twist this time. [102:58] Can everyone strike their favorite pose. [103:02] Gemini take a photo of this amazing audience, [103:04] but turn it into a cartoon and add a big blimp [103:07] in the sky that says Google I/O 2026 on it in fun colors. [103:11] Everyone strike a pose. [103:14] OK, if this goes well, everyone go bananas in the audience [103:20] because Nano banana sunglasses is just awesome. [103:23] And just in a few seconds, you'll [103:25] even see that seamless preview on her watch. [103:29] Any second now. [103:31] Drum roll. [103:33] Here we go. [103:43] Thank you, thank you. [103:44] A round of applause for nishtha. [103:51] The demo worked. [103:52] Yes the future of intelligent eyewear [103:56] has never been more exciting. [103:58] Stunning designs from brands, engineering and craftsmanship [104:04] from Samsung. [104:05] Personal and proactive help from Gemini. [104:09] Customized apps and features from Google and the developer [104:12] ecosystem. [104:13] All arriving with our first glasses this fall. [104:17] So stay tuned. [104:19] Now I'm going to pass it to Dennis. [104:23] Thank you. [104:25] I'm going to pass it to Dennis to talk about the future of AI. [104:37] DEMIS HASSABIS: It's amazing to see how far we've come with glasses [104:40] and I can't wait for everyone to experience them. [104:44] Today we showed you our next generation models Gemini 3.5 [104:47] and Omni. [104:48] New coding capabilities in Antigravity 2.0, [104:52] agents in search and Gemini Spark and so much more. [104:56] It's great to see Gemini transforming [104:58] so many Google products used by billions of people every day. [105:03] All of these advances show the staggering pace of AI progress. [105:06] It's incredible. [105:08] Even for those of us who have spent our entire lives working [105:11] on this, AGI is now on the horizon [105:14] and it will be the most profound and impactful technology [105:17] ever invented. [105:19] If built right, it could propel human progress and flourishing [105:23] beyond our imaginations. [105:25] We're in a moment of immense promise, [105:27] but also enormous responsibility. [105:30] It's important that we are clear eyed [105:32] about the potential challenges, and use all the tools [105:34] at our disposal to ensure the safety of our agentic systems, [105:38] and ultimately, AGI itself. [105:42] One area of risk that has gained a lot of attention [105:44] recently is cybersecurity. [105:46] Google has invested in this area for decades, [105:49] and we're bringing our frontier capabilities and deep expertise [105:52] to help secure the world's code bases. [105:55] We have tools like our code security agent, code mender, [105:59] which automatically finds and fixes critical software [106:02] vulnerabilities. [106:04] Today, we're inviting a select group of experts [106:06] to test a new code mender API, and we'll be launching it [106:09] more broadly soon. [106:17] Stepping back, the whole reason I've worked on AI [106:20] my entire career was because I saw it as the ultimate tool [106:23] to advance science and our understanding of the world. [106:27] It's awesome to see those dreams become a reality, [106:30] with AI beginning to help scientists [106:32] in almost every field. [106:35] Building on this momentum, I'm excited to announce [106:37] Gemini for science, which brings together. [106:46] It brings together a number of powerful AI tools [106:49] to help accelerate research. [106:51] Gemini can already Assist in solving complex problems, [106:55] but our new lab's prototypes streamline [106:57] daily scientific tasks. [106:59] Whether it's staying on top of newly published papers, [107:02] transforming research goals into usable code, [107:05] or generating new hypotheses, another powerful tool [107:09] for science is simulation. [107:11] AI simulations are going to be critical for understanding [107:14] and predicting dynamic systems that [107:16] are simply too complex to model directly today. [107:20] An amazing example of this is alpha Earth foundations. [107:24] It's the closest thing we have to a digital twin of the planet [107:27] that could help address problems like deforestation and food [107:30] security. [107:32] Simulations are already proving extremely useful. [107:36] Our state of the art weather. [107:37] Next, models can predict hurricane paths faster and more [107:40] accurately than traditional systems. [107:43] Let's take a look at how weather next helped during last year's [107:47] hurricane season. [107:49] Tropical storms and hurricanes can change very quickly, [107:52] which makes them more challenging to predict [107:54] than other types of weather systems. [107:56] At Google, we developed weather next, a global weather [107:59] forecasting AI model that is also [108:00] able to predict where hurricanes are going to go [108:02] and how strong they're going to become in 2025. [108:06] Weather next predicted a category 5 hurricane [108:08] striking Jamaica three days early [108:10] with greater accuracy than previous models. [108:13] It's going to cause catastrophic, [108:16] life threatening damage. [108:23] Because of that early Warning, we [108:26] were able to give that advance notice to the public to say, [108:30] move from certain areas. [108:31] So it saved their lives. [108:34] Weather next was a really valuable tool [108:36] in helping us make these more accurate and aggressive [108:38] forecasts for Melissa. [108:40] And I think going forward, weather next [108:42] and other AI models will become part of our routine forecast [108:45] toolkit here at the Hurricane Center. [108:56] In the future, we will be able to simulate [108:58] even more complex emergent systems, [109:00] perhaps even virtual cells. [109:03] Our biological models like AlphaFold and alpha genome, [109:06] have already become standard research tools [109:08] used by millions of scientists around the world [109:11] to make important advances in their fields. [109:14] I like to call this science at digital speed, [109:18] both in terms of the speed of the solution [109:20] and its dissemination to the researchers who [109:22] can make use of it. [109:24] I've always believed the number one application of AI [109:27] should be to improve human health. [109:30] At Isomorphic Labs, we're modeling molecular interactions [109:34] to massively accelerate the development [109:36] of new medicines supported by leading industry partners. [109:40] We're now in preclinical stage with multiple projects, [109:44] including potential treatments for immune disorders and cancer. [109:48] Our mission. [109:55] Our mission is to reimagine the drug discovery process [109:58] with the goal of one day solving all disease, something [110:03] that would have seemed impossible just a few years ago. [110:06] But I truly believe it's now within reach. [110:10] Google's cutting edge research and products [110:12] will help unlock AGI's incredible potential [110:15] for the benefit of the entire world. [110:18] When we look back at this time, I [110:20] think we all realize that we were [110:22] standing in the foothills of the singularity. [110:25] It will be a profound moment for humanity. [110:28] This technology will be a force multiplier for human ingenuity [110:32] and usher in a new golden age of scientific discovery [110:35] and progress, improving the lives of everyone, everywhere. [110:41] We look forward to building the future with all of you. [110:44] Thank you and enjoy the rest of Google I/O.