Quick Overview
This video is an episode of The Code Report by Fireship covering the release of Google's flagship frontier model, Gemini 4 Argon. Host Jeff Delaney evaluates Google's newest architecture in response to competing releases from Anthropic and OpenAI, analyzing benchmark data, pricing, and live coding performance.
Key Points
- 1.Google announced Gemini 4 Argon, a new frontier mixture of experts foundation model trained from scratch on Tensor Processing Units using JAX.
- 2.Argon leads benchmarks such as DeepSWE v1.1 and the GDP-weighted Vals Index, but trails Claude Opus 5.5 on Terminal-bench 4.0 and GPT-6 Astra on FrontierSWE v2.
- 3.Google launched Argon with introductory pricing of 2 dollars per million input tokens and 10 dollars per million output tokens.
- 4.Access controversies arose after Google removed Gemini 3.1 Pro from Antigravity and restricted Argon to the 99.99 dollar monthly Google AI Ultra plan.
- 5.In practical Three.js coding tests, Gemini 4 Argon produced an unexpected anatomical anomaly on a 3D horse model, highlighting alignment and safety discrepancies.
Summary
The video begins by noting recent tech updates before turning to Google's unveiling of Gemini 4 Argon, framed as a major bid to reclaim ground in the frontier AI race after Gemini 3.5 Pro failed to hit internal release targets. Built as a brand-new sparse mixture of experts transformer rather than a fine-tune, Argon was trained from scratch over six months on Google Tensor Processing Units using the JAX library and ML Pathways, carrying a knowledge cutoff date of March 2026.
Argon was initially unveiled through the Fairwind Program to select cybersecurity enterprises without standard guardrails, before expanding with safety filters designed to sharply reduce refusal rates. Google positioned Argon at an aggressive introductory price through November 30th of 2 dollars per million input tokens and 10 dollars per million output tokens, before shifting to standard rates of 4 dollars and 20 dollars respectively. Benchmarks show Argon leading in several key categories, posting 77.9 percent on DeepSWE v1.1, 68.9 percent on the Vals Index, 19.6 percent on Harvey's Legal Agent Benchmark, and 91.7 percent on LVBench. However, competitor models still lead in critical areas: Claude Opus 5.5 beat Argon 66.4 to 57.4 on Terminal-bench 4.0, GPT-6 Astra outperformed it by more than ten points on FrontierSWE v2, and Argon dropped to ninth place on Arena AI's WebDev Code Arena.
The commercial rollout caused controversy among developers. Google removed Gemini 3.1 Pro from its Antigravity development environment and restricted Argon to subscribers of the Google AI Ultra plan, which costs 99.99 dollars per month. This decision left standard Pro plan subscribers locked out until lower quotas were later granted.
Hands-on testing pitted Argon against Claude Opus 5.5 and GPT-6 Astra across complex creative tasks. When tasked with creating promotional material for a hypothetical Horse Tinder app, Argon generated a direct AI video file, whereas Claude produced scripted motion graphics by outputting executable code. In a Three.js 3D rendering evaluation requiring an interactive horse, Claude and GPT-6 Astra produced realistic interactive 3D models with controls. Argon built a polished user interface, but rendered a rudimentary low-poly horse featuring an erect, conical copulatory organ. This bizarre failure raised immediate questions about Google's alignment and safety enforcement.
The video wraps up with a profile of Macroscope, an AI code review tool built to manage high volumes of agent-generated pull requests. Macroscope uses custom abstract syntax tree walkers across languages like Go, Rust, and TypeScript to verify PR safety and automatically approves roughly 40 percent of low-risk pull requests without human intervention.
Gemini 4 Argon Architecture and Benchmarks
Google introduced Gemini 4 Argon as a new foundation model built from scratch over six months using Google TPUs, JAX, and ML Pathways rather than fine-tuning older releases. The sparse mixture of experts model features a March 2026 knowledge cutoff and achieved top scores on long-horizon software engineering benchmarks like DeepSWE v1.1 with 77.9 percent, the Vals Index with 68.9 percent, and video analysis on LVBench with 91.7 percent. However, it placed second to Claude Opus 5.5 on Terminal-bench 4.0 and lost to GPT-6 Astra on FrontierSWE v2 by over ten points.
Rollout Strategy and Developer Backlash
Originally previewed via the Fairwind Program to select cybersecurity partners, Argon was released with promotional API pricing through November 30th of 2 dollars per million input tokens and 10 dollars per million output. Despite attractive token economics, Google drew criticism from developers when it removed Gemini 3.1 Pro from its Antigravity IDE and locked Argon access behind the 99.99 dollar per month Google AI Ultra subscription tier, temporarily leaving Pro subscribers without access to an elite coding model.
Practical Multi-Modal and Code Generation Testing
Practical experiments evaluated Argon against Claude Opus 5.5 and GPT-6 Astra across multimedia generation and 3D web design. When prompted to generate video assets for an application called Horse Tinder, Argon outputted raw synthetic video directly, whereas Claude generated animated motion graphics by compiling clean code. In a Three.js coding challenge requiring a realistic interactive 3D horse, Argon produced a functional user interface but generated a low-poly horse model with an anatomically inappropriate, prominent phallic shape that raised safety and alignment questions.
Automated Code Review with Macroscope
The broadcast concludes with a review of sponsor Macroscope, a developer platform designed to handle the growing influx of pull requests generated by automated coding agents. Macroscope uses abstract syntax tree walkers to evaluate code changes across popular programming languages, enforcing team rules and assessing blast radius. By screening pull requests, the platform auto-approves approximately 40 percent of low-risk code submissions without requiring manual human review.
The Bottom Line
The episode demonstrates that while Gemini 4 Argon delivers competitive benchmark scores and aggressive token pricing, it remains uneven in real-world coding compared to top offerings from Anthropic and OpenAI. Google's rollout strategy stirred developer frustration by gating the model behind an expensive tier and stripping access to older tools. The unusual Three.js generation bug also leaves open whether Google's relaxed refusal policies have created new safety and alignment blind spots.
FAQ
What is Gemini 4 Argon and what are its key technical specifications?
Gemini 4 Argon is Google's frontier AI model built from scratch as a sparse mixture of experts transformer, trained over six months on Tensor Processing Units using JAX and ML Pathways, featuring a knowledge cutoff of March 2026.
How much does Gemini 4 Argon API usage cost during the introductory period?
During its introductory promotional window through November 30th, Argon costs 2 dollars per million input tokens and 10 dollars per million output tokens, after which pricing rises to 4 dollars for inputs and 20 dollars for outputs.
Which benchmarks did Gemini 4 Argon win and lose against competing AI models?
Gemini 4 Argon placed first on DeepSWE v1.1 with 77.9 percent, the Vals Index with 68.9 percent, and LVBench with 91.7 percent. However, it lost to Claude Opus 5.5 on Terminal-bench 4.0 and was beaten by GPT-6 Astra by over ten points on FrontierSWE v2.
Why were developers frustrated with Google's Antigravity platform rollout of Argon?
Developers were angered because Google removed access to Gemini 3.1 Pro from Antigravity and initially locked Argon behind the Google AI Ultra subscription, which costs 99.99 dollars per month.
What happened when Gemini 4 Argon was tested on generating a Three.js horse?
While Gemini 4 Argon produced a polished user interface, the 3D horse itself was low-poly and unexpectedly rendered with an anatomically explicit, large erect phallus, revealing safety and alignment flaws.
Worth watching for
Software engineers and AI developers looking for a technical breakdown of Google's Gemini 4 Argon, its benchmark rankings against competitor models, and its real-world coding performance.
- gemini-4-argon
- google-deepmind
- large-language-models
- claude-opus
- gpt-6
- ai-benchmarks