Google started rolling out Gemini 4 Argon on Tuesday, TheNextWeb and Axios reported, but the first people to get it aren’t ordinary users. They’re cybersecurity defenders enrolled in Fairwind, a voluntary US government program that gives vetted security teams early access to frontier AI models.
Early access comes without the usual guardrails
Fairwind participants get Argon without its standard safety restrictions, letting them push the model harder than a consumer ever could. That approach already paid off once: security firm Wiz used the unrestricted model to find a critical vulnerability in healthcare software that earlier frontier models had missed entirely. It’s a notable real world result, not just a benchmark claim, and it’s the kind of case Google is clearly using to justify handing out an unguarded model before the public gets a filtered one.
What the benchmarks actually show
Argon scored 77.9% on DeepSWE v1.1, a benchmark built around real software engineering tasks, and tied for first place on CWE-bench, which tests how well a model fixes genuine security flaws, at 68%. Google says Argon beats OpenAI’s GPT-6 Astra on multiple coding and knowledge-work evaluations, a direct shot at the model that’s been setting the pace since its own launch this week. Tulsee Doshi, who heads Gemini products at Google DeepMind, called Argon “a well-rounded model that has frontier capabilities” across several domains, language that signals Google wants this seen as an all-purpose flagship rather than a narrow security tool.
Price drops while the context window explodes
Argon’s API pricing lands at $2 per million input tokens and $10 per million output tokens. The bigger jump is the context window, now 1 million tokens, up from roughly 64,000 on the previous Gemini generation, a sixteen-fold increase that lets the model hold far more code or documentation in a single session without losing track of earlier context. For developers who hit the ceiling on long codebases or lengthy legal documents, that change matters more day to day than any single benchmark score.
Not everyone inside Google is convinced yet
Paying API customers and Google AI Ultra subscribers get access next, with broader availability expected only after more testing. Bloomberg has reported that some Google employees are skeptical the benchmark gains will translate cleanly into everyday use, a reminder that strong scores on DeepSWE or CWE-bench don’t automatically mean the model handles messier, real world requests just as well. Argon also arrives roughly a year after Gemini 3, following reports that an earlier June 2026 target slipped amid internal morale issues at DeepMind, context that makes Tuesday’s limited rollout look more like a cautious restart than a victory lap.
Source: TheNextWeb and Axios