The one OpenAI announcement that can actually make you money...

Fireship

AI-assisted summary of The one OpenAI announcement that can actually make you money... — original source. Check the original for full context.

Source published: 2026-10-01

Key ideas

OpenAI launched 'Dots,' autonomous personal agents powered by the upcoming GPT-6 Astra model priced at $100 per month, originating from the viral open-source ClaudeBot project. (0:00)
The new Pro 500 tier ($500/month) introduces UltraFast access at 300 tokens per second, but steep API rates of $60 per million input and $300 per million output tokens drew audience groans. (2:23)
OpenAI introduced the Decisions API and the Luna model to replace slow, error-prone JSON parsing loops with sub-second, deterministic enum classifications. (3:14)
'Sign in with ChatGPT' enables developers to authenticate third-party apps and shift token inference expenses directly onto users' personal ChatGPT subscriptions. (4:18)

Detailed notes

## Background Published on YouTube by the developer channel Fireship on October 1, 2026, this video provides a rapid-fire breakdown of the key announcements from OpenAI DevDay 2026. According to the description, the video covers everything developers missed at the event while highlighting Fastino's open-weight alternative models.

## TL;DR The video recaps OpenAI DevDay 2026, highlighting the rollout of autonomous personal agents named Dots alongside high-end pricing tiers like the $500 Pro 500 plan and high-cost UltraFast execution. OpenAI also introduced the Decisions API, powered by the lightweight model Luna to streamline deterministic enum classification and eliminate JSON parsing loops, as well as native clones of productivity apps like Notion and Google Docs. Finally, the video spotlights 'Sign in with ChatGPT' as a genuine monetization breakthrough that offloads inference costs onto end-user subscriptions, contrasted against cost-saving open-weight decision models from Fastino.

## Atmosphere & style Fast-paced, satirical, and irreverent, the video crackles with Fireship's signature deadpan humor and Silicon Valley skepticism. The delivery treats multi-million-dollar AI keynote spectacles as fodder for biting jokes about domain blunders, exorbitant token pricing, and the corporate habit of cloning partner startups. Watching the video feels like an entertaining, cynical tour of developer reality that punctures tech marketing hype while offering pragmatic engineering takeaways.

## Bulletin - [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) OpenAI launched 'Dots,' autonomous personal agents powered by the upcoming GPT-6 Astra model priced at $100 per month, originating from the viral open-source ClaudeBot project. - [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) The new Pro 500 tier ($500/month) introduces UltraFast access at 300 tokens per second, but steep API rates of $60 per million input and $300 per million output tokens drew audience groans. - [03:14](https://www.youtube.com/watch?v=No-JPdFvYWU&t=194s) OpenAI introduced the Decisions API and the Luna model to replace slow, error-prone JSON parsing loops with sub-second, deterministic enum classifications. - [04:18](https://www.youtube.com/watch?v=No-JPdFvYWU&t=258s) 'Sign in with ChatGPT' enables developers to authenticate third-party apps and shift token inference expenses directly onto users' personal ChatGPT subscriptions.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) OpenAI introduced Dots, a commercial personal agent platform powered by GPT-6 Astra and priced at $100 per month to compete with Meta's Muse. - [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) The Dots architecture was built on ClaudeBot, the fastest-growing repository in GitHub history created by Peter Steinberger before his hire by OpenAI. - [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) During the keynote demo, an agent named Alfred autonomously sunsetted an inventory API, generated patches, ran test suites, and opened three unreviewed pull requests. - [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) OpenAI faced branding snags as dots.com belongs to a shuttered clothing retailer, while dot.com is owned by xAI and redirects directly to Grok. - [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) The Pro 500 tier provides 25 times the quota of ChatGPT Plus alongside access to UltraFast, which streams GPT-6 Astra at 300 tokens per second. - [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) UltraFast API compute carries premium pricing of $60 per million input tokens and $300 per million output tokens. - [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) OpenAI quietly cut the monthly quota of its reintroduced $200 subscription tier in half. - [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) GPT-6.1 Soul offers near-Astra capabilities at reduced pricing of $2 per million input and $10 per million output tokens. - [03:14](https://www.youtube.com/watch?v=No-JPdFvYWU&t=194s) The Decisions API utilizes a specialized lightweight model called Luna to make constrained selections in milliseconds, eliminating JSON formatting errors and retries. - [03:14](https://www.youtube.com/watch?v=No-JPdFvYWU&t=194s) OpenAI escalated competition with existing software ecosystems by announcing Pages (a Google Docs clone) and Space (a Notion competitor). - [04:18](https://www.youtube.com/watch?v=No-JPdFvYWU&t=258s) 'Sign in with ChatGPT' provides a monetization workaround by letting developers offload AI token consumption costs directly onto end users' subscriptions. - [04:18](https://www.youtube.com/watch?v=No-JPdFvYWU&t=258s) Open-weight alternatives like Fastino's GLiNER 2.5 Decide and Glide allow developers to run low-latency decision routing locally without vendor lock-in.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) OpenAI Dev Day 2026: Meet Dots and the Personal Agent Race

The stage for OpenAI DevDay 2026 was set by escalating corporate rivalry. Just days prior, Sam Altman reportedly watched in frustration from his bunker as Mark Zuckerberg revealed Muse—a free personal assistant that essentially acts as a personal spy, training on everything users feed it in exchange for handling their shopping. Not to be outdone, Altman took the stage in San Francisco to present OpenAI's answer: a harmless, cute personal assistant priced at $100 per month designed to clear AI-generated email clutter and rewrite back ends in "whatever language DHH thinks is fast that day."

The lineage of this agent traces back to January 2026, when retired Austrian developer Peter Steinberger launched ClaudeBot. Pitched as an assistant for people who want to look productive without doing any work, the project quickly became the fastest-growing repository in GitHub history, prompting OpenAI to hire Steinberger immediately. While Meta repackaged the concept into Muse, OpenAI launched its commercial offering under the name **Dots**—where an individual "Dot" represents an autonomous agent powered by the upcoming GPT-6 Astra.

During the keynote, Altman demonstrated his personal Dot, dubbed "Alfred," tasking it with sunsetting an old inventory API. The agent navigated the task using modern developer workflows: tracing dependencies, patching integrations, running test suites, and ultimately opening three pull requests under the CEO's name—leaving developers forced to review code that no human eye had ever seen. The subsequent live presentation by Holly Lee served as a reminder that live software demos remain precarious, even in an era dominated by AI generation. The rollout also triggered immediate domain mockery across the web: `dots.com` remains tied to a women's clothing retailer that shuttered in 2014, while `dot.com` had already been acquired by xAI, actively redirecting visitors straight to the Grok landing page.

- **ClaudeBot origins**: Created by Peter Steinberger in January 2026; set the record as the fastest-growing repo on GitHub before Steinberger was hired by OpenAI. - **Dots pricing & model**: $100 per month; individual agents run on GPT-6 Astra. - **Domain mishaps**: `dots.com` belongs to an inactive apparel brand; `dot.com` is owned by xAI and redirects to Grok.

---

[02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) New Pro Pricing Tiers, UltraFast, and GPT-6.1 Soul

OpenAI restructured its developer tiers to accommodate the heavy compute required by its new models, introducing the Pro 500 tier at $500 per month. This subscription grants developers 25 times the usage capacity of ChatGPT Plus, alongside access to "UltraFast," an accelerated execution tier delivering GPT-6 Astra at 300 tokens per second. API pricing for UltraFast sparked audible moans from the keynote audience when Altman revealed rates of $60 per million input tokens and $300 per million output tokens. Altman shrugged off the audience's reaction with a deadpan defense: "You know what? It's worth it."

For developers on existing tiers, the news was less favorable. OpenAI quietly reintroduced the $200 per month subscription plan, but heavily slashed its allocated usage in half. To cater to developers working on leaner budgets, OpenAI unveiled GPT-6.1 Soul, marketed as offering intelligence near the level of GPT-6 Astra at a fraction of the operating cost.

- **Pro 500 Plan**: $500/month; provides 25× the quota of ChatGPT Plus and access to UltraFast. - **UltraFast Performance & API Rates**: GPT-6 Astra running at 300 tokens per second; priced at $60 per 1M input tokens and $300 per 1M output tokens. - **$200 Plan**: Restocked but with its monthly usage limits cut by 50%. - **GPT-6.1 Soul Rates**: $2 per 1M input tokens and $10 per 1M output tokens (matching the pricing of the previously released GPT-6 Soul).

---

[03:14](https://www.youtube.com/watch?v=No-JPdFvYWU&t=194s) The Decisions API and the Rise of OpenAI Clones

To solve one of the most frustrating inefficiencies in building agentic systems—generating long text outputs only to extract structured fields through error-prone parsing and retry loops—Altman introduced the **Decisions API**. Instead of querying a heavy frontier model to draft text and parse JSON, developers pass a constrained, predefined list of allowed answers to a lightweight model named Luna, which selects an option in a fraction of a second.

The video illustrates this architectural shift with a tongue-in-cheek example: an app analyzing user messages to classify whether text is useful for blackmail. In the conventional approach, an application calls a heavy LLM, waits several seconds while the model works through its safety guardrails, and hopes the returning JSON matches the schema. With the Decisions API, the classification is constrained to fixed enums up front; Luna resolves the choice in a couple hundred milliseconds without retry loops, because the model cannot physically output an out-of-spec value. The concept closely mirrors Jev, a tool recently launched by Type-Safe AI (a startup established by an ex-OpenAI researcher). OpenAI's DevDay announcements also took aim at productivity software, revealing Pages (a Google Docs competitor), Space (a Notion alternative), and an internal Slack agent reminiscent of Claude Tag.

```mermaid flowchart TD accTitle: Traditional LLM Parsing vs OpenAI Decisions API accDescr: Flow diagram contrasting heavy LLM JSON extraction with the lightweight Decisions API Luna architecture.

subgraph Traditional["Traditional LLM Workflow"] A["User Input"] --> B["Frontier LLM Processing"] B --> C["Extract JSON & Run Validation"] C -->|"Invalid JSON"| D["Retry Loop / Re-prompt"] C -->|"Valid"| E["Application Logic"] end

subgraph DecisionsAPI["Decisions API Workflow"] F["User Input + Fixed Enum Options"] --> G["Luna Lightweight Model"] G -->|"Deterministic Selection (Sub-second)"| H["Application Logic"] end ```

- **Luna model role**: Designed specifically for fast enum and classification matching. - **Latency & Reliability**: Operates within a few hundred milliseconds and eliminates JSON parsing retry logic. - **New productivity tool clones**: Pages (Google Docs clone), Space (Notion clone), and an autonomous token-heavy Slack bot.

---

[04:18](https://www.youtube.com/watch?v=No-JPdFvYWU&t=258s) "Sign in with ChatGPT" and Open-Weight Decision Models

The single announcement from DevDay with direct commercial viability for developers is **Sign in with ChatGPT**. This authentication protocol allows end users to sign into third-party web apps using their existing ChatGPT credentials, offloading the token costs directly onto the user's account rather than forcing the host platform to foot the API bill. The feature debuted alongside 16 launch partners, including Devin, Notion, Vercel, and OpenClaw—though partnering comes with the ever-present threat that OpenAI might eventually build a native clone of any product that gets too popular.

Because relying entirely on OpenAI's proprietary cloud introduces vendor lock-in, developers seeking low-latency classification logic often turn toward open-weight alternatives. While proprietary tools like Jev keep weights closed inside their cloud platform, Fastino offers an open-weight alternative with **GLiNER 2.5 Decide**, which can run classification pipelines locally up to eight times faster. The underlying GLiNER architecture has logged over 50 million downloads since 2024 and integrates cleanly into developer tooling like OpenCode to route prompts, extract paths, and parse function names. Fastino also introduced **Glide**, an upgraded model benchmarked to surpass Jev across intent classification, fact-checking, and hallucination detection, with developers reporting up to 36-fold cost reductions when running browser automation agents.

- **Sign in with ChatGPT**: Allows third-party apps to let users burn their personal ChatGPT quotas; debuted with 16 launch partners (Devin, Notion, Vercel, OpenClaw). - **Fastino GLiNER 2.5 Decide**: Open-weight decision model running locally up to 8× faster than closed cloud alternatives. - **Fastino Glide**: Specialized model surpassing Jev across intent routing, fact checks, and hallucination reduction. - **Cost optimization**: Browser agent workloads reported to run up to 36× cheaper using Fastino's architecture.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) whatever language DHH thinks is fast that day.

> [00:00](https://www.youtube.com/watch?v=No-JPdFvYWU&t=0s) people who want to look productive without doing any work

> [02:23](https://www.youtube.com/watch?v=No-JPdFvYWU&t=143s) You know what? It's worth it. — Sam Altman

## Mentioned OpenAI, OpenAI DevDay 2026, Sam Altman, Mark Zuckerberg, Muse, DHH, Peter Steinberger, ClaudeBot, GitHub, Dots, GPT-6 Astra, Holly Lee, xAI, Grok, Pro 500, ChatGPT Plus, UltraFast, GPT-6.1 Soul, GPT-6 Soul, Decisions API, Luna, Jev, Type-Safe AI, Pages, Google Docs, Space, Notion, Claude Tag, Slack, Sign in with ChatGPT, Devin, Vercel, OpenClaw, Fastino, GLiNER 2.5 Decide, OpenCode, Glide

## Video description Try the faster, open-weight version of Jev at https://fastino.ai/. Use code: FIRESHIPFIVE to get free credits.

In this video we'll cover everything you missed at OpenAI Dev Day 2026. Let's dive in.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=No-JPdFvYWU)

中文简报

OpenAI 这项真正能帮你赚钱的发布…… 该视频回顾了 OpenAI DevDay 2026，重点介绍了名为“Dots”的自主个人智能体的推出，以及 500 美元的 Pro 500 方案和高成本的 UltraFast 执行模式等高端定价梯队。OpenAI 还推出了由轻量级模型 Luna 驱动的 Decisions API，以简化确定性枚举分类并消除 JSON 解析循环，同时还推出了类似 Notion 和 Google Docs 的生产力应用原生克隆产品。最后，视频重点介绍了“Sign in with ChatGPT”，将其视为一项真正的变现突破，可将推理成本转嫁至最终用户的订阅中，并将其与 Fastino 节约成本的开放权重决策模型进行了对比。 OpenAI 推出了自主个人智能体“Dots”，该产品由即将推出的 GPT-6 Astra 模型驱动，定价为每月 100 美元，其灵感源于风靡一时的开源 ClaudeBot 项目。 (0:00) 新的 Pro 500 档位（每月 500 美元）提供了每秒 300 个 token 的 UltraFast 访问，但每百万输入 token 60 美元、每百万输出 token 300 美元的高昂 API 定价引来了现场观众的叹息。 (2:23) OpenAI 推出了 Decisions API 和 Luna 模型，用亚秒级的确定性枚举分类取代缓慢且易错的 JSON 解析循环。 (3:14) “Sign in with ChatGPT”使开发者能够对第三方应用进行身份验证，并将 token 推理开销直接转移到用户的个人 ChatGPT 订阅上。 (4:18)

Reading summary

At OpenAI DevDay 2026, the company introduced autonomous agents, tiered compute subscriptions, and constrained decision APIs, highlighted by a "Sign in with ChatGPT" authentication model that allows third-party applications to offload API token expenses directly to end users. Offloading Developer Costs via ChatGPT Login (4:18) According to Fireship, the single DevDay announcement offering immediate commercial utility for software builders is "Sign in with ChatGPT." This authentication protocol lets third-party web applications authenticate users with their personal OpenAI credentials, shifting underlying token consumption costs onto the end user's account rather than requiring the host platform to foot the API bill. The feature debuted across 16 launch partners, including Devin, Notion, Vercel, and OpenClaw. However, the report highlights a strategic tension for developers: integrating closely with OpenAI carries the risk of platform redundancy, as OpenAI has repeatedly launched native competitors against breakout software categories. Autonomous Agents and Dots Infrastructure (0:00) OpenAI introduced "Dots," a commercial autonomous personal agent system priced at $100 per month and powered by the upcoming GPT-6 Astra model. The underlying concept originated in January 2026 with ClaudeBot, an open-source tool created by Austrian developer Peter Steinberger that rapidly became GitHub's fastest-growing repository before OpenAI recruited him. During the keynote, Sam Altman demonstrated an agent named "Alfred" that sunset an inventory API by tracing dependencies, running tests, and opening three automated pull requests without human review. The launch also triggered web domain issues: dots.com remains tied to a defunct apparel retailer, while xAI purchased dot.com to redirect users directly to Grok. High-End Compute Tiers and Usage Rebalancing (2:23) To accommodate compute-intensive models, OpenAI unveiled a Pro 500 subscription tier at $500 per month, offering 25 times the capacity of ChatGPT Plus and access to an "UltraFast" execution tier. Running GPT-6 Astra at 300 tokens per second, UltraFast API pricing was set at $60 per million input tokens and $300 per million output tokens—rates Altman defended despite audience pushback. Lower-tier users faced tighter limits. OpenAI quietly reintroduced its $200 per month plan while reducing its allocated monthly usage by 50%. For developers needing lower-cost inference, the company introduced GPT-6.1 Soul at $2 per million input tokens and $10 per million output tokens, matching the rates of the earlier GPT-6 Soul. Constrained Decision APIs and Fastino Alternatives (3:14) To eliminate the latency and retry loops common when asking large models to generate and validate structured JSON, OpenAI introduced the Decisions API. Powered by a lightweight model called Luna, the system restricts outputs to predetermined enum values to deliver deterministic classifications in a fraction of a second. This design mirrors Jev, a tool launched by Type-Safe AI, alongside OpenAI's new application clones such as Pages and Space. To counter cloud vendor lock-in and API fees, developers are turning toward open-weight architectures. Fastino's GLiNER 2.5 Decide runs classification tasks locally up to eight times faster than closed cloud alternatives, while its newer Glide model reportedly outperforms Jev across intent classification, fact-checking, and hallucination reduction, reducing costs up to 36-fold in browser automation pipelines. While OpenAI is enabling developers to pass compute expenses to end users and streamlining classification through constrained models, high API fees and the persistent threat of native platform replication continue to push builders toward local, open-weight architectures.

Source quotes

[0:00] whatever language DHH thinks is fast that day. [0:00] people who want to look productive without doing any work [2:23] Sam Altman: You know what? It's worth it.

Original transcript

[0:00] Last week, Sam Altman sat frustrated in [0:02] his bunker as he watched Zuck unveil [0:04] Muse, a free personal assistant-inspired [0:06] spy that trains on everything you tell [0:08] it in exchange for doing your shopping. [0:10] Not to be outdone, Sam got on a stage of [0:12] his own in San Francisco just a few days [0:14] ago and announced that for $100 per [0:17] month, they'll give you another cute and [0:19] totally harmless personal assistant that [0:21] can respond to all of those AI-generated [0:23] emails in your inbox and rewrite your [0:25] back end in whatever language DHH thinks [0:27] is fast that day. In today's video, [0:29] we'll break down everything else you [0:30] missed at OpenAI Dev Day, including a [0:32] model that's almost as smart as GPT-6 [0:35] Astra, an API that looks suspiciously [0:37] familiar, [0:38] >> I'm not Sam Altman. [0:39] >> and a new subscription tier that quietly [0:41] castrated the $200 per month plan. It is [0:43] October 1st, 2026 and you're watching [0:46] The Code Report. Back in January, a [0:48] bored and recently retired Austrian [0:50] developer named Peter Steinberger [0:52] released ClaudeBot, a first-of-its-kind [0:54] AI personal assistant for people who [0:56] like to appear productive without the [0:58] hassle of actually doing the work. It [0:59] quickly became the fastest-growing repo [1:01] in GitHub history, which was enough to [1:03] get him hired by OpenAI to continue to [1:05] work on the idea. Then, just last week, [1:08] all his hard work finally paid off when [1:10] Zuck took the concept and stuffed it [1:11] into Muse. And just a few days ago, his [1:14] own employer did the same thing, but now [1:16] with a $100 per month price tag that [1:18] they're calling Dots, with a singular [1:20] dot being a personal agent powered by [1:22] GPT-6 Astra. So, it's basically just a [1:25] bot with the B flipped around. You just [1:26] give it a name, then its cute, squishy [1:28] little avatar will get to work ruining [1:30] the internet as we know it. In their [1:32] keynote demo, Sam told his dot, Alfred, [1:34] to remove an old inventory API before it [1:36] got shut down. So, Alfred then followed [1:39] modern software best practices of [1:41] tracing dependencies, updating [1:42] integrations, running tests, and [1:45] spamming the repo with three PRs from [1:47] the CEO that some poor dev will feel [1:49] obligated to review because they know [1:51] it'll be the first time human eyes have [1:53] ever looked at that code. Then, Holly [1:55] Lee came out to prove that Dots work [1:57] outside of a pre-recorded video and [1:59] instead just proved that live demos are [2:01] still hard even in the age of slop. But [2:03] my favorite part of dots is when the [2:05] internet realized that dots.com is owned [2:07] by some woman's clothing brand that died [2:10] in 2014 and dot.com is owned by xAI and [2:14] now redirects to grokbots landing page. [2:16] It's the best domain prank I've seen [2:18] since whoever got their hands on [2:19] nextjs.dev and remix.dev a few years [2:22] back. But if dot sounds like your slop [2:24] cannon of choice, you'll need a pro [2:25] subscription which open AI revealed some [2:28] new pricing for. The new pro 500 plan [2:30] costs $500 per month and gets you 25 [2:33] times the usual usage of chat GPT plus [2:36] along with access to ultrafast which is [2:38] a new speed tier that runs Astra at 300 [2:41] tokens per second. In the API, ultrafast [2:43] costs $60 per million tokens in and 300 [2:46] out. And when Sam said those words out [2:48] loud, the audience of grown men moaned. [2:51] >> You know what? It's worth it. [2:52] >> Which apparently isn't the first time [2:54] he's caused that reaction. Meanwhile, [2:55] the $200 plan went back on sale but had [2:58] its usage chopped in half. Luckily, if [3:00] you're broke, they announced GPT 6.1 [3:02] soul which they're describing as near [3:04] Astra intelligence for a fifth of the [3:06] price. You get $2 per million tokens in [3:09] and $10 out which is the exact same [3:11] price of GPT 6 soul which launched a [3:13] week ago. Then Sam announced the new [3:15] decisions API where the idea is that [3:18] instead of asking a big model to write a [3:20] paragraph and then parsing the JSON out [3:22] of it with a retry loop, you hand their [3:24] tiny Luna model a fixed list of answers [3:26] and it picks one in a fraction of a [3:28] second. For example, imagine you're [3:29] building a dating app for gay donkeys. [3:31] As a way to fund the venture, every [3:33] message needs to be categorized by how [3:34] useful it would be for blackmail. [3:36] Historically, you'd send each message to [3:38] a full language model, wait a few [3:40] seconds while it pretends to be ethical, [3:41] then hope the JSON comes back in the [3:43] right shape. With a decision model, you [3:45] give it a finite set of answers up [3:47] front. It returns one of them in a few [3:48] hundred milliseconds and since it [3:50] physically can't answer anything outside [3:52] your list, there's nothing to parse or [3:54] retry. And if all this sounds familiar, [3:56] it should because just 2 weeks ago, we [3:58] released a video about a company called [4:00] Type-Safe AI. They're founded by an [4:01] OpenAI researcher who just released Jev [4:04] that does the exact same thing. But they [4:06] didn't stop there. Later in the event, [4:07] they announced a Google Docs clone [4:09] called Pages, a Notion clone called [4:11] Space, and a Slack bot that works a lot [4:14] like Claude Tag showing what it looks [4:16] like when you have infinite tokens. But [4:18] the one thing in the keynote that could [4:19] actually make you money is sign in with [4:21] ChatGPT, which lets your users log into [4:23] your app with their ChatGPT account and [4:26] burn their own tokens instead of yours. [4:27] It launched with 16 partners including [4:30] Devin, Notion, Vercel, and OpenClaw and [4:32] comes with an implicit understanding [4:34] that if you get too popular, they'll [4:35] just clone your app, too. Which is why [4:37] if you need to build a decision model [4:39] into your own app, you might want to own [4:41] the weights yourself. And you can do [4:42] that with Fastly Labs, the sponsor of [4:44] today's video. Jev has been going viral [4:47] for giving developers fast type [4:49] decisions, but it's closed weight and [4:51] only runs on the company's cloud. But [4:52] Fastly's Gliner 2.5 Decide is an open [4:56] weight model that makes those same calls [4:58] locally and up to eight times faster. [5:00] This so you can train your own Jev-like [5:02] models and own the weights. It comes [5:04] from the Gliner family, which has been [5:05] open source since 2024 with over 50 [5:08] million downloads. I added their skill [5:10] to OpenCode in about 5 minutes so my [5:12] agent can route each developer request [5:14] to a coding model or internal tool and [5:17] pull out the repo file path and function [5:19] names. Other developers have used it to [5:20] run a browser agent 36 times cheaper [5:23] than Jev. And Fastly just released a [5:25] brand new model called Glide that beats [5:27] Jev on accuracy across intent routing, [5:29] fact checking, and hallucination checks [5:32] by a pretty wide margin. You can host [5:33] these models anywhere, but Fastly's [5:35] inference API is the quickest way to [5:37] replace your frontier model. Try it out [5:39] for free with our code at the link [5:41] below. This has been the Code Report. [5:42] Thanks for watching and I will see you [5:44] in the next one.