Google announced Gemini 4 Argon on 30 September 2026 and began rolling it out to trusted cybersecurity defenders through its Fairwind Program, VentureBeat reported in its account of the launch. CNBC described Argon as Google’s most advanced AI model yet. Access is starting with defenders rather than a broad public release, while the company takes part in a voluntary US government process for evaluating models before wider availability.
Key points
- Initial access is going to trusted cybersecurity defenders through Google’s Fairwind Program.
- Google is participating in a voluntary US government pre-release model access process.
- Argon leads or ties on 13 of the 18 benchmarks in Google’s disclosed comparison.
- The model’s stated output limit rises to 1 million tokens from 64,000.
Fairwind gives defenders the first access
The initial group is narrow for a model intended eventually for much wider use. Google says it is beginning with trusted cyber defenders in Fairwind and participating in the US government’s voluntary pre-release model access process, according to VentureBeat. The company plans broader availability for developers, enterprises and consumers, beginning with paid API customers and Google AI Ultra subscribers. Its stated timing for that step is “as soon as possible”.
The choice reflects the model’s intended defensive use and the risks of making those capabilities widely available. Tulsee Doshi, Google’s Gemini model product lead, told CNBC that the limited start lets the company put a model trained for cyber defence in defenders’ hands while gaining confidence in the rollout. Google is working to expand safeguards against misuse and prompt injection before a public launch. A prompt-injection attack attempts to make a model follow hostile instructions placed in material it is asked to examine.
Trusted defenders and Google’s internal teams receive Argon without cyber guardrails, Techzine reported. It said the model can find, check and patch vulnerabilities, while Google is strengthening protections against misuse involving biological, chemical, radiological and nuclear threats, as well as indirect prompt injection, for wider access. That makes the first release a controlled deployment of capabilities Google also wants to make useful for security work.
The announcement follows a period in which Google kept releasing smaller Flash models while competitors advanced their frontier models, Axios reported. An earlier AI Affairs report described Gemini 4 in post-training ahead of release. On 29 September, Google chief executive Sundar Pichai signed a voluntary AI safety accord with President Donald Trump and other technology executives at the White House, CNBC reported.
DeepSWE puts Argon’s coding score at 77.9%
Google’s disclosed comparison covers 18 benchmarks, and Argon leads outright on 12 and ties for the highest score on one, VentureBeat reported. On DeepSWE v1.1, an evaluation of longer software-engineering tasks, Google’s table gives Argon 77.9%, against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On CWE-bench v1, which tests vulnerability remediation, Argon and GPT-6 Astra each score 68% in the same table.
The comparison also gives GPT-6 Astra a higher score on FrontierSWE v2, at 65.5% against Argon’s 55.0%, and Claude Opus 5.5 a higher score on Terminal-bench 4.0, at 66.4% against 57.4%. These are results on the named tests in Google’s disclosed set. The differing scores matter to companies choosing a model for a particular kind of work: a lead on one software task does not carry automatically to another.
Google says Argon can produce up to 1 million output tokens, up from a previous limit of 64,000. Tokens are the units used to count a model’s generated text, so the output limit is a ceiling on how much it can write in one response, rather than a measure of whether that writing is correct. Techzine lists a price of $2 per million input tokens and $10 per million output tokens.
Reviewing a long piece of code could involve asking for a more extended written response before it reaches that stated output ceiling. Whether the response would identify the right changes would still depend on the task: Google’s disclosed software benchmarks give Argon different scores for different kinds of work.
Google applies Argon to data-centre memory
Google is also using Argon inside its own operations. The company says it has applied the model to memory use in its data centres, freeing hundreds of terabytes without buying additional hardware. That is an account of an internal deployment rather than a benchmark score: it concerns a change to working infrastructure, with the saving reported by the company operating it.
