Google ships Gemini 4 Argon with a 1M-token output window
Google DeepMind's Gemini page presents Gemini 4 Argon as "our next era of frontier intelligence, built to solve your hardest problems". It is stated to be available today for Google AI Pro and Ultra subscribers, which the page calls Google's flagship subscription tier for power users, developers and creators. Two application programs open alongside it: academic and independent scientists can apply for the highest usage tier in Google Antigravity free of charge for 6 months (subject to eligibility, capacity and selection), and cybersecurity defenders can apply to join the Fairwind Program.
The headline technical feature is a 1M-token output window. Google says that headroom lets Argon finish 100K+ token reasoning trajectories and produce step-by-step mathematical proofs, quantum error-correction decoders and complete computational physics solvers in one go. Other listed uses are generating whole multi-file applications or refactoring large codebases in a single pass, translating monolithic legacy files and database schemas end to end, and producing compliance audits, technical specifications and large structured datasets in a single turn.
The showcase example is Chrome. Google says Argon helped the Chrome team separate core webpage handling from a ~6 million line legacy folder of browser features inside Chromium's 40 million-line codebase. Piping multiple clean code changes at a time, Argon helped finish an architectural refactor across hundreds of files, completing months of work in weeks. Google says this will let live webpages move between tabs, side panels and floating windows without reloading, and will eliminate an entire category of browser crashes. That is stated as a future effect.
On cybersecurity, the page says Argon ties for #1 on CWE-Bench (68%) for vulnerability remediation and is Google's most resilient model yet against indirect prompt injections. The Fairwind Program is expanding to hundreds of additional partners today. Wiz, which had early access, is using Argon for its "Scan for Good" initiative, which protects critical public infrastructure for free; Google says Wiz used it to protect critical healthcare infrastructure and that it safeguarded hundreds of thousands of patient records.
Most of the performance evidence is partner testimonials, none attributed to a named individual. Box's evaluation reported a 14% improvement in accuracy over Gemini 3.1 Pro. A 3D spatial reconstruction partner reported cutting runtime nearly 100x and token usage by 90%. Another partner said long-running tasks completed successfully 100% of the time in its evals. Kepler said Gemini 4 matched its analyst-built reference models on more than 99% of figures in financial analysis benchmarks. A partner running an autonomous AI researcher called Qualia reported evals completed 1.5x faster than peer models at a quarter of the cost per task. A medical evaluation on over 200 complex synthetic longitudinal charts found the highest recall and ICD-10 code specificity among the models tested. Roboflow said Argon ranked first in its object detection evaluations and that it has fully integrated the model. LexisNexis reported strong preliminary results on legal knowledge, drafting and analysis, and another legal partner said the model surpassed its human expert's answer key in one evaluation. Poe said Gemini 4 gives creators a powerful foundation for building tools.
The safety section lists four measures. Argon is designed to refuse harmful requests while preserving legitimate dual-use scientific research, per Google's Frontier Safety Framework. Google is deploying misalignment mitigations that monitor Argon's chain-of-thought and actions and stop execution when necessary. It is also hardening sandboxed environments by isolating and sealing them before high-risk training or evaluations begin. The prompt-injection resilience claim sits alongside these.
Key facts
- Gemini 4 Argon is stated to be available today for Google AI Pro and Ultra subscribers; academic and independent scientists can apply for 6 months of the highest Google Antigravity usage tier free, subject to eligibility, capacity and selection.
- Argon has a 1M-token output window, which Google says lets it finish 100K+ token reasoning trajectories and generate or refactor whole codebases in a single pass.
- Argon ties for #1 on CWE-Bench at 68% for vulnerability remediation, and the Fairwind Program for cyber defenders is expanding to hundreds of additional partners.
- Google says Argon helped the Chrome team complete an architectural refactor across hundreds of files in weeks rather than months, in a ~6 million line legacy folder within Chromium's 40 million-line codebase.
- Much of the performance evidence is partner testimonials, including a 14% accuracy gain over Gemini 3.1 Pro in Box's evaluation and more than 99% figure matches in Kepler's financial benchmarks.
Why it matters
Google positions Argon as its next frontier model, and the page leans on long-horizon work: very long single-pass outputs and sustained engineering tasks. The Chrome refactor, months of work completed in weeks across hundreds of files in a ~6 million line folder, is the page's concrete example of that. The 1M-token output window is the most specific product feature stated. The cyber results matter too: Argon ties for #1 on CWE-Bench at 68%, and the Fairwind Program is widening to hundreds more defender partners.
Who it affects
Google AI Pro and Ultra subscribers get the model first. Software engineers are the page's main audience, with Google saying Argon is quickly becoming the preferred model for engineers on long-horizon software engineering across Google and early access partners. Security teams and cybersecurity defenders are another target through the Fairwind Program. Academic and independent scientists can apply for the Antigravity tier. Partners named in testimonials span legal (LexisNexis), finance (Kepler), computer vision (Roboflow), document management (Box) and consumer creation (Poe).
How to use it
Argon is available today to Google AI Pro and Ultra subscribers, and the page offers a "Try Gemini 4 Argon" entry point. Academic and independent scientists can apply for the highest usage tier in Google Antigravity free for 6 months, subject to eligibility, capacity and selection. Cybersecurity defenders can apply to join the Fairwind Program. The page states no API availability or API token prices for Gemini 4 Argon.
How solid is it
This is Google's own launch page, so the claims are first-party. Only one benchmark score is given explicitly: 68% on CWE-Bench, tying for #1. The "Performance" section has only a heading and no benchmark table. The other figures come from partner testimonials, none attributed to a named individual, and they come from the partners' own evaluations. The 14% gain over Gemini 3.1 Pro is Box's evaluation, not a Google benchmark. The page does not say which models are meant by "peer models" or "every model we tested".
Risks and caveats
Several claims are forward-looking or unverified. The Chrome benefits are stated as what the work "will unlock", and the page does not say whether the refactor has shipped to users. The prompt-injection claim is that Argon is Google's "most resilient model yet", a comparison against its own earlier models rather than an absolute guarantee. Google describes monitoring of chain-of-thought and actions plus sandbox hardening, which indicates it treats misalignment and high-risk evaluation as live concerns. Model size, architecture and training details are not given. The page does not say whether Argon is available to free-tier users or through the API at launch.
“Gemini 4 represents a step change in what AI can accomplish in real-world science.”
— Unnamed partner testimonial on Google DeepMind's Gemini 4 Argon page