Gemini 4 Argon stays with vetted cyber defenders

Gemini 4 Argon is Google's new frontier model, and the October 1 coverage made the restriction the story: the weights are not a public download, and the first outside users are a vetted set of cyber defenders in the Fairwind Program. Google's own post, dated September 30 and bylined by Koray Kavukcuoglu, says the company is in the U.S. government's voluntary pre-release access process and will gather tester feedback before a broader release.

TL;DR - Google announced the model for Fairwind cyber defenders and its own teams, not for a general consumer or developer launch. - The introductory API price in the company post is $2 per million input tokens and $10 per million output tokens, later $4 and $20. - SecurityWeek and The Guardian both treated the limited rollout as a safety choice, not as a finished public product.

What actually happened

The primary account is Google's post, Introducing Gemini 4. It says the model is built for long-horizon work in software engineering, knowledge work such as legal and finance, and cybersecurity defense. It also says output length is being expanded to 1 million tokens, from a previous 64,000. Those figures come from Google. They are claims about Gemini 4 Argon, not independent benchmark reproductions.

Google lists internal examples. It says the model beat a published baseline by 40% on a quantum subroutine's spacetime resources. It says a team of agents identified memory optimizations that freed more than 300 TiB, with an estimated 500 TiB to 1 PiB of further savings. It describes C and C++ to Rust migrations, including a libgav1 result it says runs 2.7 times faster than an existing Rust port. On named evals, Google says the model set a new state of the art on DeepSWE v1.1 at 77.9%, ranked first on Zapier's AutomationBench at 51.3%, and scored 91.7% on LVBench. On CWE-bench v1 it says the model ties for first at 68%.

What the other pages add

SecurityWeek, in a piece by Eduard Kovacs timestamped 3:52 a.m. Eastern on October 1 (10:52 a.m. MSK), adds names Google's post did not spell out in the text opened here: it says the 68% tie is with OpenAI's GPT-6 Astra and xAI's Grok 4.7. That attribution belongs to SecurityWeek's reading of the launch, so it should sit beside Google's "ties for first" line rather than replace it. The same article says Fairwind launched in early September with Gemini 3.8 Flash Cyber and CodeMender, and that the program then had more than 650 partners. It also notes that the hospital-software finding is unnamed, and that the announcement does not say whether the issue has been fixed.

The Guardian, via Agence France-Presse, published at 12:47 p.m. Eastern on October 1 (7:47 p.m. MSK). It quotes the "phased approach" line and reports that Google is giving the U.S. government early access. It places the rollout next to Anthropic's restricted Claude Mythos Preview, and it says the announcement came a day after a White House meeting where tech executives signed a voluntary risk accord. Those political details are in the AFP report. They are not in the Google post opened for this draft.

On price, only the company post gives numbers. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. After the introductory period, Google says the price becomes $4 per million input tokens and $20 per million output tokens. No second outlet opened here republished a different price, so there is no conflicting quote to average. The wider release, Google says, starts with paid API customers and Google AI Ultra subscribers, and only after more guardrail work.

For trusted defenders and internal teams, Google says it will release the model without cyber guardrails so they can use full defensive capability. For everyone else, the post describes refusals on cyber and chemical, biological, radiological, and nuclear misuse, activation monitoring, prompt-injection hardening, and a check that can stop a run if chain-of-thought or actions look misaligned. Wiz, which Google says is using the model in Scan for Good, is the named early tester behind the hospital-software example. Google says earlier frontier models missed that exposure. It does not name the software.

Readers who want a sense of how other labs have been shipping agent tooling can look at Grok 4.7's write-up and the OpenClaw Enterprise note. Those are separate products. They are not substitutes for a model that Google itself has not put in general availability.

What to do as a reader

Treat Gemini 4 Argon as a restricted research and defense rollout, and keep the name Gemini 4 Argon tied to Fairwind until Google names a public date, a console, and a model ID you can call. If a social post offers "early access" keys, downloads, or a third-party endpoint, that is not what either Google or SecurityWeek described. Read the benchmark lines as the company's own scores unless a lab posts a reproduction. The 68% CWE-bench tie is a leaderboard claim. SecurityWeek names the tied models. Google's post, in the version opened here, does not.

If you follow cyber defense work, the useful questions are narrow. Who is actually inside Fairwind? Did Wiz or the hospital vendor publish the vulnerability? Did the price on the developer page match the $2 and $10 introductory pair, or the later $4 and $20 pair? Those checks do not require buying a token or treating a demo as a finished product.

This is a news reading of two opened reports plus the company post. It is not a recommendation to use, avoid, or pay for any model.