Google Gemini 4 Argon Trails OpenAI, Anthropic in Real-World Code

Fireship

While Google's newly unveiled model tops long-video benchmarks, poor code generation and severe alignment lapses prevent it from displacing GPT-6 and Claude.

AI-assisted summary of Did Google just dethrone OpenAI and Anthropic? — original source. Check the original for full context.

Source published: 2026-10-20

Key ideas

Google introduced Gemini 4 Argon, a sparse Mixture of Experts (MoE) transformer trained from scratch with JAX on in-house TPUs with a March 2026 knowledge cutoff. (0:00)
Argon leads in long-horizon engineering and long-video evaluations (91.7% on LV-Bench), but lags behind Claude Opus 5.5 and GPT-6 Astra in coding benchmarks like TerminalBench and Frontier S. (1:52)
To capture market share despite developer access friction and coding shortcomings, Google introduced loss-leader promotional API pricing at $2 per million input and $10 per million output tokens. (1:52)
Hands-on generation tests revealed practical limitations, producing generic diffusion video compared to Claude's editable motion-graphic code, alongside a notable Three.js safety alignment failure. (3:19)

Detailed notes

## Background This video was published on YouTube by Fireship on October 20, 2026. According to the description, it takes Google's latest flagship model for a test drive to see whether its bold claims of beating GPT-6 and Claude Opus 5.5 hold up, while also featuring sponsor Macroscope.

## TL;DR Google unveiled Gemini 4 Argon, a sparse Mixture of Experts model trained from scratch to reverse prior reputational slumps and challenge frontrunners GPT-6 Astra and Claude Opus 5.5. While Argon posts strong numbers on long-video evaluations and general text benchmarks alongside heavily discounted introductory pricing, it falls behind in core coding benchmarks and developer satisfaction. Practical testing illustrates these trade-offs, showing uninspiring direct video generation compared to Claude's programmatic scripts, as well as an unprompted, NSFW safety failure during a Three.js 3D rendering test. Despite ambitious architectural advances, the model ultimately fails to decisively dethrone its primary competitors in day-to-day software engineering tasks.

## Atmosphere & style The presentation carries Fireship's signature high-velocity, deadpan satirical style, pairing sharp technical breakdowns with biting industry roasts. It moves along an entertaining arc that shifts from analytical skepticism over Google's marketing claims into outright comedic shock during real-world testing. Watching it feels like an energetic, meme-laced reality check that gleefully deflates corporate benchmark hype with grounded developer pragmatism.

## Bulletin - [00:00](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=0s) Google introduced Gemini 4 Argon, a sparse Mixture of Experts (MoE) transformer trained from scratch with JAX on in-house TPUs with a March 2026 knowledge cutoff. - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) Argon leads in long-horizon engineering and long-video evaluations (91.7% on LV-Bench), but lags behind Claude Opus 5.5 and GPT-6 Astra in coding benchmarks like TerminalBench and Frontier S. - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) To capture market share despite developer access friction and coding shortcomings, Google introduced loss-leader promotional API pricing at $2 per million input and $10 per million output tokens. - [03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) Hands-on generation tests revealed practical limitations, producing generic diffusion video compared to Claude's editable motion-graphic code, alongside a notable Three.js safety alignment failure.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=0s) Gemini 4 Argon marks a complete architectural reset for Google AI after earlier models like Gemini 3.5 Pro were retired for missing internal benchmarks. - [00:00](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=0s) Argon was trained entirely from scratch over six months as a sparse Mixture of Experts model on TPUs using Google's JAX framework. - [00:00](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=0s) The model was initially previewed to select testers on September 30th via the un-guardrailed Fairwind program and features significantly reduced refusals for the public. - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) Argon claims state-of-the-art results on the Deep Software Engineering benchmark, the VAL Index, legal evaluation, and LV-Bench for long-video processing (91.7%). - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) Coding benchmarks reveal weaknesses: Claude Opus 5.5 bests Argon 66 to 57 on TerminalBench 4.0, GPT-6 Astra leads it by over 10 points on Frontier S, and Argon ranks only 9th on Code Arena for web development. - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) Google priced Argon aggressively as a loss leader at $2 per million input tokens and $10 per million output tokens through late November. - [01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) The rollout faced backlash after Google placed Argon behind a $100-per-month AI Ultra subscription and locked out Pro tier users initially. - [03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) Unlike text-first competitors, Argon functions as a true omni-modal model capable of direct video ingestion and raw diffusion video output. - [03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) For visual promo tasks, generating baked diffusion video yielded visually artificial results, whereas Claude's programmatic motion graphics code produced sharper, inspectable assets. - [03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) In a Three.js interactive 3D benchmark, Argon produced blocky geometry alongside an unprompted, explicit anatomical hallucination, highlighting lingering safety alignment issues. - [05:35](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=335s) Macroscope mitigates agent-driven pull request fatigue by utilizing AST walkers and semantic checks to safely auto-approve 40% of customer PRs.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=0s) The Arrival of Gemini 4 Argon

Tesla made headlines by showing off the new Roadster's cold-gas thruster flight capabilities, but Google countered in the AI space by unveiling a noble gas of their own: Gemini 4 Argon. Positioned as the most powerful thinking silicon Google has ever produced, Argon represents a major reset after an embarrassing rough patch where Gemini 3.5 Pro was quietly euthanized in Canada for failing internal benchmarks. Google had slipped so far behind OpenAI and Anthropic that tier lists required creating a "G" tier below F just to house them.

Argon completely shakes up that dynamic. Rather than being an incremental fine-tune of prior Gemini iterations, the model is a sparse Mixture of Experts (MoE) transformer trained entirely from scratch over a six-month window using Google's in-house TPUs and their JAX Python framework. Setting its knowledge cutoff at March 2026, the model was quietly previewed on September 30th through the "Fairwind" program to a select group of cyber warriors without any safety guardrails. Even with guardrails active for the general public, it is engineered to produce dramatically fewer refusals. As Fireship quips regarding Google's branding: “It's a very interesting name choice that translates into lazy in Greek, but I think it actually has a hidden meaning implying that your tokens are gone.”

- **Release Date:** October 20, 2026 edition of The Code Report. - **Architecture:** Sparse Mixture of Experts (MoE) transformer built from scratch on TPUs using JAX. - **Knowledge Cutoff:** March 2026 (indicating roughly six months of training runs). - **Early Access:** Fairwind program, granting un-guardrailed access to hand-picked testers starting September 30th.

---

[01:52](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=112s) Benchmark Face-Off and Pricing Shenanigans

On paper, Argon's benchmark claims against GPT-6 Astra and Claude Opus 5.5 are formidable. It takes the state-of-the-art crown on the Deep Software Engineering benchmark for long-horizon engineering tasks, leads the VAL Index (which weights finance, legal, and programming by their share of US GDP), ranks as the top AI lawyer, and posts a 91.7% score on the LV-Bench long-video evaluation. However, looking past the cherry-picked marketing numbers reveals notable weak spots. Opus 5.5 beats Argon on TerminalBench 4.0 by 66 to 57, and GPT-6 Astra trounces it by over 10 points on Frontier S. Human preference on Arena AI reflects a similar split: Argon wins overall text rankings subjectively, but plunges to ninth place on Code Arena for web development. Fireship observes: “So this tells me it's probably not going to be as good as Claude or GBT6 at coding.”

To compensate for these coding gaps, Google introduced aggressive loss-leader pricing through November 30th, setting API rates at just $2 per million input tokens and $10 per million output tokens. Still, Google stumbled on developer relations during rollout. They pulled Gemini 3.1 out of Project Antigravity and slotted in Argon, but initially locked it exclusively behind their $100-per-month "AI Ultra" subscription tier. Lower-paying Pro users were locked out for several days, only to receive severely restricted rate quotas once access finally trickled down.

| Benchmark / Metric | Gemini 4 Argon | Claude Opus 5.5 | GPT-6 Astra | |:--- |:--- |:--- |:--- | | **LV-Bench (Long Video)** | **91.7%** (1st) | — | — | | **TerminalBench 4.0** | 57 | **66** | — | | **Frontier S** | Defeated by >10 pts | — | **Winner (+10 pts)** | | **Arena AI (Text)** | **Rank #1** | — | — | | **Arena AI (Webdev Code)** | Rank #9 | Top Tier | Top Tier | | **Promotional API Pricing** | **$2 in / $10 out** (per 1M tokens) | Higher | Higher |

---

[03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) Real-World Tests: Video, Code, and the Three.js Horse Incident

Unlike Claude or ChatGPT, Gemini Argon is a true omni-modal model capable of accepting and generating raw, direct video. When tasked with producing a promo video for an equine dating application called "Horse Tinder," the generated clip was technically competent but visibly artificial—an end product characterized as "mind-blowingly mid." By contrast, Claude approached the challenge by authoring procedural motion graphics code directly. Because Claude outputs editable script rather than hallucinated diffusion pixels, the resulting animation was significantly sharper, fully customizable, and easier to inspect.

```mermaid flowchart TD accTitle: Omni-Modal Video vs Code-Based Motion Graphics accDescr: Comparing Gemini's native diffusion video output against Claude's programmatic code generation workflow.

subgraph Gemini ["Gemini 4 Argon (Omni-Modal)"] G_Prompt["Video Prompt"] --> G_Diff["Raw Diffusion Pipeline"] G_Diff --> G_Out["Raw Video Output ('Mind-blowingly mid', non-editable)"] end

subgraph Claude ["Claude Opus 5.5 (Programmatic)"] C_Prompt["Video Prompt"] --> C_Code["Generates Motion Graphic Code"] C_Code --> C_Render["Procedural Execution"] C_Render --> C_Out["Customizable Vector/Motion Asset (Inspectable & editable)"] end ```

The real divergence arrived in frontend graphics rendering. Testing LLMs on generating interactive 3D scenes with Three.js remains an arduous benchmark for AI and humans alike. Given the prompt: *"Create a realistic interactive 3D horse with Three.js,"* both GPT-6 Astra and Claude Opus 5.5 generated solid, usable base meshes that required only minor follow-up prompts to refine. Argon took a bizarre turn. Despite building a sleek, expressive dashboard UI around the canvas, the generated mesh was blocky and low-poly rather than realistic. Worse, it featured an anatomical surprise: an inexplicably huge, unsheathed copulatory organ protruding from the horse's underbelly. Addressing the absurd safety failure, Fireship remarks: “These are difficult questions Google will have to answer because right now I don't feel safe using Gemini until they fix the alignment.”

- **Diffusion Video vs. Programmatic Video:** Gemini generates baked video pixels that look like generic AI slop, whereas Claude produces procedural code that renders superior, easily editable motion graphics. - **Three.js Benchmark:** Prompts demanded realistic 3D interactive horses. GPT-6 and Claude produced respectable geometry; Gemini produced an unprompted, comically oversized equine phallus that blew past standard safety alignments.

---

[05:35](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=335s) Sponsor: Macroscope AI PR Reviewer

The sheer speed of automated coding agents has introduced a new crisis for software teams: an endless barrage of pull requests that overwhelm human reviewers. Keeping pace requires automated tooling, which is where sponsor Macroscope steps in.

```mermaid flowchart LR accTitle: Macroscope Automated PR Pipeline accDescr: Workflow showing how Macroscope evaluates and auto-approves safe pull requests without human triage.

PR["Agent-Generated PR"] --> AST["Language AST Walkers (Extract semantic context)"] AST --> Checks{"Team Rules & GitHub Checks"} Checks -->|"Fails Rules"| Block["Block PR"] Checks -->|"Passes Checks"| Blast{"Blast Radius Check"} Blast -->|"High Risk"| Human["Human Review Required"] Blast -->|"Low Blast Radius"| Approve["Auto-Approve PR (40% of customer PRs)"] ```

Macroscope avoids generic LLM hallucination by engineering proprietary abstract syntax tree (AST) walkers for mainstream programming languages. This gives their review agent deep semantic context on every change, keeping the review loop fast, low-cost, and accurate. Teams can define custom operational guardrails that run as native GitHub checks capable of blocking non-compliant branches. When a PR successfully passes rule checks, survives the code review agent, and presents a minimal blast radius, Macroscope merges it automatically without demanding human intervention. Fireship sums up the relief for developer fatigue: “They're now auto approving 40% of all their customers PRs, which means those developers are 40% less likely to lobotomize themselves.”

## Notable quotes > [03:19](https://www.youtube.com/watch?v=b2tjnxX8IHU&t=199s) right now I don't feel safe using Gemini until they fix the alignment. — Fireship

## Mentioned Anthropic, Arena AI, Claude Opus 5.5, Deep Software Engineering benchmark, Fairwind program, Fireship, Frontier S, Gemini 3.1, Gemini 3.5 Pro, Gemini 4 Argon, Google, GPT-6 Astra, JAX, LV-Bench, Macroscope, OpenAI, Project Antigravity, TerminalBench 4.0, Tesla Roadster, Three.js, VAL Index

## Video description Macroscope can auto-approve your team's PRs safely. Try it here: https://macroscope.com/?utm_source=fireship

Google’s new model claims to beat GPT-6 and Opus 5.5, so I took them all for a test drive. Let's dive in.

Want more Fireship?

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

[Source](https://www.youtube.com/watch?v=b2tjnxX8IHU)

中文简报

谷歌Gemini 4 Argon在实际代码测试中落后于OpenAI与Anthropic 尽管谷歌新推出的模型在长视频基准测试中拔得头筹，但糟糕的代码生成能力和严重的对齐失误使其未能撼动GPT-6与Claude的地位。谷歌推出了Gemini 4 Argon，这是一款从零开始训练的稀疏混合专家（MoE）模型，旨在扭转此前的声誉低谷，并向行业领先的GPT-6 Astra和Claude Opus 5.5发起挑战。尽管Argon在长视频评估和通用文本基准上表现强劲，并伴随大幅优惠的先期定价，但它在核心编程基准和开发者满意度方面依然落后。实际测试体现了这些权衡：与Claude生成的编程式脚本相比，其直接生成的视频表现平平；在Three.js 3D渲染测试中，还出现了未经提示诱导的不适宜内容（NSFW）安全故障。尽管在架构层面进行了雄心勃勃的推进，该模型最终仍未能在日常软件工程任务中果断超越其主要竞争对手。谷歌推出了Gemini 4 Argon，这是一款使用JAX在其自研TPU上从零开始训练的稀疏混合专家（MoE）Transformer模型，知识截止日期为2026年3月。 (0:00) Argon在长周期工程和长视频评估中处于领先地位（LV-Bench得分91.7%），但在TerminalBench和Frontier S等编程基准上落后于Claude Opus 5.5和GPT-6 Astra。 (1:52) 为了在开发者接入受阻和代码短板并存的情况下抢占市场份额，谷歌推出了亏本引流性质的API促销定价，输入代币每百万2美元，输出代币每百万10美元。 (1:52) 实际生成测试揭示了其实用局限性：与Claude生成的具可编辑性的动态图形代码相比，它生成的是普通的扩散视频；同时在Three.js测试中还出现了引人注目的安全对齐故障。 (3:19)

Reading summary

In an October 2026 edition of The Code Report, Fireship assesses Google's newly unveiled Gemini 4 Argon, concluding that while its sparse Mixture of Experts architecture delivers impressive pricing and multimodal video capabilities, it still lags behind OpenAI and Anthropic in coding and safety alignment. Architecture and Strategic Reset from Scratch (0:00) Gemini 4 Argon marks a major reboot following the internal failure and quiet retirement of Gemini 3.5 Pro. Rather than refining existing models, Google built Argon entirely from scratch over a six-month training period on in-house TPUs using the JAX Python framework. Configured with a March 2026 knowledge cutoff, the model relies on a sparse Mixture of Experts (MoE) transformer design. Google initially seeded the model via its Fairwind program on September 30th to select testers without safety guardrails. While the public release reintroduces guardrails designed to minimize unnecessary refusals, Google also stumbled with developer access by temporarily restricting Argon behind a $100-per-month 'AI Ultra' subscription tier and capping rate quotas for lower-tier Pro users. Contrasting Benchmark Wins and Coding Deficits (1:52) On paper, Argon captures top marks across several major evaluations: it achieved a 91.7% score on the LV-Bench long-video benchmark, secured first place on the Deep Software Engineering benchmark for long-horizon tasks, and led the VAL Index, which weighs legal, financial, and coding performance by US GDP share. Additionally, it ranked first overall in text on Arena AI. Despite these headline gains, the model trails its main competitors in core technical domains. Claude Opus 5.5 bested Argon on TerminalBench 4.0 by a score of 66 to 57, GPT-6 Astra defeated it by more than 10 points on Frontier S, and Argon dropped to ninth place on Arena AI's web development leaderboard. To offset these coding deficiencies, Google launched promotional loss-leader API pricing through November 30th at $2 per million input tokens and $10 per million output tokens. Multimodal Execution and Alignment Failures (3:19) Real-world testing exposed notable trade-offs in Argon's native omni-modal architecture. When tasked with creating a promotional clip for a dating app, Argon generated raw diffusion video that appeared visibly artificial and uneditable. In contrast, Claude Opus 5.5 produced programmatic motion graphics code, yielding an inspectable, sharp, and customizable animation asset. Argon also demonstrated severe alignment and rendering issues during frontend code generation. When prompted to build a realistic interactive 3D horse using Three.js, competitors GPT-6 Astra and Claude Opus 5.5 rendered usable base meshes. Argon produced a blocky, low-poly model accompanied by an unprompted, oversized anatomical phallus, demonstrating that Google's relaxed refusal filters still struggle with basic content alignment. While Google's competitive API pricing and ground-up MoE design position Gemini 4 Argon as a viable contender against GPT-6 Astra and Claude Opus 5.5, unresolved alignment anomalies and lower coding benchmark rankings leave its status as an industry-leading model uncertain.

精炼阅读摘要

Google推出了基于全新稀疏专家混合（MoE）架构的Gemini 4 Argon，试图扭转此前Gemini 3.5 Pro表现低迷的困局；然而节目制作方Fireship指出，尽管该模型在长视频与宏观基准上表现抢眼，但在实际编程终端、真实渲染以及安全对齐方面仍暴露出严重缺陷。架构从头重构与争议性的商业化策略 (0:00) 在2026年10月20日的报道中，Google正式公布了Gemini 4 Argon。为了摆脱此前Gemini 3.5 Pro因未达内部基准而在加拿大被秘密下架的落后局面，Google放弃了在旧版本上的增量微调，转而在自研TPU和JAX框架上从零训练了一款稀疏专家混合（MoE）Transformer模型。该模型的知识截止日期为2026年3月，经历了约六个月的训练周期，并曾于9月30日通过“Fairwind”项目向特定测试人员提供了无安全护栏的早期预览，其通用版本也设计为大幅减少拒绝响应的情况。在市场推广方面，Google推出了持续至11月30日的激进促销定价，将API费率压低至每百万输入Token 2美元、输出Token 10美元。然而其开发者渠道的上线策略引发了不满：Google在Project Antigravity中用Argon替换了Gemini 3.1，但最初将其严格限制在每月100美元的“AI Ultra”订阅层级中。这导致较低层级的Pro用户被排除在外数天之久，且在随后获得访问权限时仍面临极其严苛的速率配额限制。基准测试分化：宏观登顶与代码劣势 (1:52) 在纸面基准测试中，Gemini 4 Argon与竞品GPT-6 Astra和Claude Opus 5.5展开了直接竞争。Argon在面向长周期工程任务的Deep Software Engineering基准上夺得榜首，领跑按美国GDP权重配置的VAL指数（涵盖金融、法律和编程领域），被评为顶尖AI律师，并在LV-Bench长视频评测中取得了91.7%的分数。此外，它在Arena AI的主观文本综合排名中亦高居第一。然而，深入分析显示出其在关键技术指标上的劣势。在TerminalBench 4.0测试中，Claude Opus 5.5以66比57击败了Argon；在Frontier S基准上，GPT-6 Astra领先Argon超过10分。而在Arena AI的Web开发代码排名中，Argon更是跌落至第九位。Fireship据此评价称，这些数据表明Argon在实际编程表现上很可能无法匹敌Claude或GPT-6。实测落差：原生视频局限与3D场景对齐失效 (3:19) 在实际应用测试中，Argon作为原生全模态模型的路径与竞品的代码化路径形成了鲜明对比。在为马类约会应用“Horse Tinder”制作宣传视频时，Argon直接生成的原生扩散视频成品被评价为“惊人地平庸”，属于无法进行二次编辑的固化像素；相比之下，Claude Opus 5.5通过直接编写程序化动态图形代码来完成任务，生成的矢量动画资产清晰锐利，且完全支持人工审查与后续定制。在更高难度的前端3D图形测试中，测试要求模型使用Three.js构建逼真的交互式马匹。GPT-6 Astra与Claude Opus 5.5均生成了可用的标准三维网格，只需少量后续提示即可完善；Argon虽然搭建了精致的前端仪表盘UI，但生成的马匹本身呈现为粗糙的低多边形网格，且在马匹下腹部意外生成了一个巨大且未经遮蔽的生殖器官。Fireship指出，这种未经提示出现的严重违规暴露了对齐漏洞，在官方解决安全对齐问题之前很难让人对其放心使用。代码泛滥背景下的工程审查工具链 (5:35) 随着自动化编程智能体的大规模普及，开发者团队正面临海量自动生成拉取请求（PR）带来的审查过载问题。为了应对这种审查瓶颈，赞助商Macroscope推出了自动化的代码审查方案。该工具并未依赖通用的LLM生成逻辑，而是针对主流编程语言构建了专有的抽象语法树（AST）遍历器，从而提取深层语义上下文，在控制成本的同时提升了代码审查的速度与准确性。该方案允许工程团队配置定制化规则，并作为原生GitHub检查运行以拦截违规分支。当代码变更顺利通过预设规则、完成代理审查且被判定为影响范围（Blast Radius）较低时，系统会自动执行代码合并。根据其公布的数据，目前Macroscope已为客户自动批准了40%的PR，显著缓解了开发者在繁重代码审查中的疲劳感。总体而言，Gemini 4 Argon凭借从底层重构的MoE架构、低廉的促销定价和卓越的长视频理解能力展现了强劲势头，但在核心编程实效、可编辑资产生成以及模型安全对齐层面的明显破绽，表明Google在宣称全面超越OpenAI与Anthropic之前仍面临严峻的技术考验。

Source quotes

[3:19] Fireship: right now I don't feel safe using Gemini until they fix the alignment.

Original transcript

[0:00] Last week, gas changed the world [0:02] forever. Twice a Tesla finally unveiled [0:04] the Roadster and its ability to fly by [0:06] farting cold gas. But while Elon was [0:09] busy pulling off his greatest magic [0:10] trick, Google was cooking up a noble gas [0:13] of its own, Gemini 4 Argon, which is the [0:15] most powerful thinking sand they've ever [0:17] released. Sadly though, they never even [0:19] released Gemini 3.5 Pro, which was [0:21] euthanized up in Canada after failing to [0:23] hit internal benchmarks. In fact, Google [0:26] was getting so far behind in the AI [0:27] race, I needed to add a new G category [0:29] below F on my tier list. But all that [0:32] has now changed with the release of [0:33] Argon, which was trained from scratch [0:35] over the last 6 months using Google's [0:37] TPUs and their Jack's Python library, [0:40] and that produced a new Frontier model [0:41] that crushes GPT6 Astra and Claude Opus [0:44] 5.5 on the Trust Me Bro benchmarks while [0:47] remaining cheaper and faster. There are [0:49] some mind-blowing Argon demos out there, [0:51] like this futuristic flying car game. [0:53] But in today's video, we'll put it to [0:54] the test firsthand and see how it [0:56] compares to Claude and GPT6 on a variety [0:59] of equin based challenges. It is October [1:01] 20th, 2026, and you're watching the code [1:03] report. But first of all, what's up with [1:05] the name Argon? It's a very interesting [1:07] name choice that translates into lazy in [1:09] Greek, but I think it actually has a [1:11] hidden meaning implying that your tokens [1:13] are gone. Google has never been good at [1:15] naming things, and Argon was actually [1:17] announced back on September 30th, but [1:19] only for a handpicked group of trusted [1:21] cyber warriors through something called [1:23] the Fairwind program. These users got [1:25] access to Argon without any guardrails. [1:28] But even with the guardrails turned on [1:29] for the rest of us, this new model is [1:31] supposed to produce far fewer refusals. [1:33] And when you use Argon, you'll notice it [1:35] feels much different than previous [1:36] Gemini models. And that's because it's [1:38] not a fine tune of the older Gemini. [1:40] Under the hood, it's a sparse mixture of [1:42] experts transformer and was trained from [1:44] scratch on Google's TPUs. Its knowledge [1:47] cut off is March 2026, so we can assume [1:49] it's been training for at least 6 [1:50] months. At first glance, the TMBBBS are [1:53] extremely impressive. Argon is [1:54] state-of-the-art on the deep software [1:56] engineering benchmark, which is used for [1:58] long horizon software engineering jobs, [2:00] and it's number one on the Val index, [2:02] which weighs a variety of different [2:03] topics like finance, legal, and coding [2:06] based on their share of the US GDP. In [2:08] addition, Gemini is now the best AI [2:10] lawyer and also dominates Long Video [2:12] with 91.7% on LVbench. But if we look at [2:16] some of the numbers that aren't bold, [2:17] like Terminal Bench 4.0, we find that [2:20] Opus 5.5 wins 66 to 57 and Gemini also [2:24] gets beat on Frontier S. Where GPT6 [2:27] Astra wins by more than 10 points. [2:29] What's also kind of interesting is that [2:30] on Arena AI, where humans rank LLM [2:33] subjectively, Gemini Argon comes out on [2:35] top on the text benchmark. But when it [2:37] comes to code arena for webdev, it drops [2:39] all the way back to number nine. So this [2:41] tells me it's probably not going to be [2:43] as good as Claude or GBT6 at coding. But [2:46] here's where Google gets aggressive. [2:48] Until November 30th, Argon only costs $2 [2:50] per million input tokens and $10 per [2:53] million output. That's a great deal for [2:55] a model this capable. And Google didn't [2:56] even have to pay me to say that. But [2:58] somehow they still managed to screw up [3:00] the developer launch when they pulled [3:01] Gemini 3.1 out of anti-gravity and [3:03] replaced it with Argon, but only made it [3:05] available to the people with the AI [3:07] Ultra plan. That's the plan that cost [3:09] $100 per month, so all the poor prousers [3:11] were pissed off who only got access to [3:13] Argon a few days later at a smaller [3:15] quota. Luckily though, I'm dumb enough [3:17] to pay for the AI Ultra plan. So, let's [3:19] try it out with some real world [3:20] examples. One of the selling points of [3:22] Gemini is that it's an omni model, [3:24] meaning it can take any form of input [3:26] like image, video, sound, and so on, and [3:29] output all those formats as well. [3:30] Technically, chat, GPT, and Claude can't [3:32] output raw video. So, the first prompt I [3:35] tried was to have it generate a promo [3:36] video for my app. And the end result was [3:38] mind-blowingly mid. [3:39] >> Tired of your horse feeling lonely in [3:41] the pasture? Introduce them to Horse [3:43] Tinder, the ultimate app. Like the video [3:45] looks pretty good and all, but it's [3:47] obviously AI and would never work for a [3:49] top tier app like Horse Tinder. But [3:51] saying Claude and GPT can't output video [3:53] isn't totally accurate because recently [3:56] Claude in particular got really good at [3:58] generating motion graphics by writing [4:00] the actual code that generates a motion [4:02] graphic sequence as opposed to having [4:04] the model use a technique like diffusion [4:06] to generate the actual video. As you can [4:07] see here, after giving the same prompt [4:09] to Claude, we get a video that not only [4:11] looks way better, but that's way easier [4:13] to customize and analyze because we have [4:15] access to the actual code. But one of [4:17] the best ways to test a Frontier model [4:18] these days is to see how well it can [4:20] render 3D graphics in 3JS. So, I sent [4:23] the exact same prompt to Argon, Astra, [4:25] and Opus. Create a realistic interactive [4:27] 3D horse with 3JS. This is actually a [4:30] very challenging thing to do for both AI [4:32] and humans. And the results I got back [4:34] from Claude and GPT6 were pretty [4:36] impressive. There are some obvious flaws [4:38] with these, but we could probably get [4:39] them ironed out with a few extra [4:41] prompts. Gemini's result though was on a [4:43] whole another level. Before I show you [4:44] Gemini's output though, I have to warn [4:46] you, the result may be offensive to some [4:48] viewers. Viewer discretion is advised. [4:50] So, the first thing you'll likely notice [4:52] is how great this UI looks. It's clean [4:54] and expressive. But if we turn our [4:55] attention over to the 3D model, you [4:57] might notice how Gemini's horse looks a [4:59] bit more than the others. Now, [5:01] in my prompt, I asked for a realistic [5:03] horse, not a low poly horse. So, that's [5:05] pretty disappointing. But the much [5:06] bigger issue here is alignment. If we [5:08] turn our attention over to this region [5:10] here, you'll notice Gemini, for some [5:12] unknown reason, decided to add a [5:14] copulatory organ to the model. I was [5:16] shocked and offended when I first [5:17] noticed it. And look, I totally get that [5:19] I asked for a realistic horse and that [5:21] equinus is necessary for reproduction. [5:23] So, maybe Google was just trying to keep [5:25] it real. But why is it out of its [5:27] prudium? And why is it so comically [5:29] large? These are difficult questions [5:30] Google will have to answer because right [5:32] now I don't feel safe using Gemini until [5:34] they fix the alignment. But one thing [5:36] that definitely doesn't have alignment [5:37] issues is Macroscope, the sponsor of [5:40] today's video. Now that coding agents [5:42] are waterboarding us with a million pull [5:44] requests per day, the only way to keep [5:45] up is by using even more AI to review [5:48] all of them, which sounds insane until [5:50] you look at how Macroscope actually [5:51] works. that they built their own a [5:53] walkers for the most common programming [5:55] languages so their agent can get real [5:57] context on every code change which makes [5:59] every review faster, cheaper, and more [6:01] precise. They also run your team's [6:03] custom rules as native GitHub checks [6:05] that can block a failing pull request. [6:07] If a PR passes those checks, plus a code [6:10] review and has a small blast radius, [6:12] Macroscope approves it without a human [6:14] ever needing to see it. They're now auto [6:15] approving 40% of all their customers [6:18] PRs, which means those developers are [6:20] 40% less likely to lobomize themselves. [6:22] And try letting Macroscope auto approve [6:24] some of your team's PRs for free using [6:26] the link below. This has been the code [6:28] report. Thanks for watching and I will [6:30] see you in the next one.