Google announces Gemini 4 Argon and gives it to vetted cyber defenders first, with its cyber guardrails switched off for them
An independent index ranks it first of 41 models. Paid API customers come later, with no date, and Google says it is in the US government's voluntary pre-release testing.
Why it matters: The model now ranked first on an independent index is going to cyber defenders before anyone else, with its cyber refusals switched off for them, while it goes through the US government's voluntary pre-release process. A lab's release order is becoming a safety control in its own right, a day after the White House accord made such controls voluntary commitments.
Google announced Gemini 4 Argon on 30 September, in a post by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google's chief AI architect. Google pitches it at long, multi-step work, and raises the output limit to 1M tokens, up from 64K.
The access plan is the unusual part. Argon “is rolling out to a set of trusted cyber defenders through our Fairwind Program”, with paid API customers and Google AI Ultra subscribers to follow; the post gives no date. For those defenders and Google's own teams, the post says “we'll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities.” To stop bad actors using it for cyber or chemical, biological, radiological and nuclear attacks, Google says the model is designed to refuse harmful requests under its Frontier Safety Framework. It also says: “We are actively engaged in the U.S. government's voluntary process for pre-release model access while we gradually expand access.”
Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, with cached input at 95% off. Google's own benchmark claims: 77.9% on DeepSWE v1.1, first place on Zapier's AutomationBench at 51.3%, and a tie for first on CWE-bench v1 at 68%.
The independent evaluator Vals AI ranks Argon first of 41 models on its Vals Index at 68.90%, ahead of Anthropic's Claude Sonnet 5.5 at 67.04%. One mismatch: Vals lists a maximum output of 262k tokens, not the 1M Google states. The announcement links no model card or system card, and two release trackers do not list Argon yet, since it has no public API.
- Confirmed Argon is rolling out first to trusted cyber defenders in Google's Fairwind Program, without cyber guardrails for them; paid API customers and Google AI Ultra subscribers follow, with no date. Google
- Confirmed Introductory price $2 / $10 per million input / output tokens, later $4 / $20; cached input 95% off; output limit 1M tokens, up from 64K. Google
- Claimed 77.9% on DeepSWE v1.1, first on Zapier's AutomationBench at 51.3%, tied first on CWE-bench v1 at 68%. Google
- Confirmed Vals Index: Argon 68.90%, first of 41 models, ahead of Claude Sonnet 5.5 at 67.04%; Vals lists maximum output as 262k tokens. Vals AI
Models & releasesSafety, security & governance
Lead story in the October 1, 2026 edition · front page