Google has begun rolling out its most advanced artificial intelligence model but access remains confined to selected cyber defenders while its performance claims face real-world testing, reported The Rundown AI.
Gemini 4 Argon entered a limited rollout on 30 September 2026 through Google's Fairwind Program, which serves vetted government, enterprise and cyber security partners. Google describes it as delivering 'frontier' performance across real-world software engineering, legal and financial work, and cyber defence while sustaining deep reasoning over complex, multistep tasks.
Ordinary Gemini users cannot yet try it. The first cohort comprises trusted cyber defenders and internal Google teams, while the company participates in the US government's voluntary process for pre-release model access, according to CNBC.
Paid application programming interface (API) customers and Google AI Ultra subscribers are due to receive access first, but Google has announced no date. Gemini product lead Tulsee Doshi called Argon an 'incredibly well-rounded' model and said the restricted start would build confidence while placing its cyber capabilities in defenders' hands quickly.
Strong Early Scores Need Wider Testing
The controlled release follows eye-catching benchmark results. Argon's high-reasoning setting placed first on Arena's 30 September text leaderboard after 4,942 votes, although Arena labelled the result preliminary, The Rundown noted.
Artificial Analysis gave Argon 53 on its Intelligence Index, matching GPT-6 Astra at its maximum reasoning setting and exceeding Gemini 3.1 Pro Preview's score of 30. Google also reported 77.9 per cent on DeepSWE v1.1, which tests long-running, real-world software engineering work.
The company claims Argon leads other evaluations covering professional tasks, long documents, charts and video. Its output ceiling has increased from 64,000 to one million tokens, giving the model greater capacity for lengthy sequences of work.
However, these tests do not establish how reliably Argon will perform for broader groups of customers. The Rundown AI said wider trials must show whether its benchmark gains translate into dependable everyday work.
Google plans introductory rates of about £1.51 ($2) per million input tokens and £7.55 ($10) per million output tokens. The prices will later double to £3.02 ($4) and £15.10 ($20), respectively, although Google has not said when the offer will end. Cached input will cost 95 per cent less than the normal input rate.
Why Google Is Holding Back Wider Access
Argon's defensive cyber abilities are central to both its appeal and the cautious rollout. Google says it can autonomously find, confirm and patch critical software vulnerabilities. Vetted defenders and internal teams will receive it without cyber guardrails, allowing them to use its full defensive capacity.
Before wider availability, Google is strengthening four areas of protection. These address malicious cyber or chemical, biological, radiological and nuclear (CBRN) requests; indirect prompt injections; behaviour that departs from a user's instructions; and weaknesses in sealed testing environments used for higher-risk evaluations.
Argon is designed to reject harmful requests, while monitoring systems can halt its actions when necessary. Google said internal and external red teams tested the safeguards through manual and automated attacks. It also claims Argon leads Gray Swan's Indirect Prompt Injection benchmark.
The release came one day after Google chief executive Sundar Pichai joined other technology leaders in signing a voluntary artificial intelligence safety agreement with US President Donald Trump, CNBC added. It also extends Google's push in the top-tier model race nearly a year after Gemini 3, following its focus on faster, lower-cost Flash models.
Early users will now test whether Argon's performance holds up outside Google's controlled environment. Until broader access begins, its headline benchmark gains remain promising but incomplete evidence of what the model can deliver at scale.
