Google announced Gemini 4 Argon on September 30, its first true frontier model release since the Gemini 3 line, and it is not shipping to the general public yet. Early access is going through something Google calls the Fairwind Program, a group of trusted cybersecurity defenders who get to use Argon for vulnerability patching before anyone else touches it. According to Google’s own announcement, Argon posts 77.9% on DeepSWE v1.1 for long-horizon software engineering and ties for first place at 68% on CWE-bench v1, a benchmark built specifically around finding and fixing vulnerabilities.
What Gemini 4 Argon Actually Does
Argon is built for long, complicated work rather than quick chat replies. Google says it targets real-world software engineering, enterprise knowledge work like legal and tax research, and cybersecurity defense, with creative writing listed as a secondary use case. The headline spec is a 1 million token output limit, a jump from the 64K ceiling on earlier Gemini models, which matters for tasks that need to hold an entire codebase or legal document set in context without losing track of earlier steps.
Google’s own testing claims Argon agents migrated more than 800,000 lines of C/C++ to Rust across enterprise systems, including parts of the Fuchsia Zircon kernel. On AutomationBench, which measures end-to-end business task completion, Argon ranks first at 51.3%. On LVBench, a long-video understanding test, it scores 91.7%.
Why Security Teams Get It Before Anyone Else
Google isn’t opening this one up the normal way. The Fairwind Program routes initial access to vetted cyber defenders first, and Google is simultaneously going through the U.S. government’s voluntary pre-release model access process rather than skipping straight to a public launch. That is a notably more cautious rollout than Google used for Gemini 3.8 Flash, which shipped a cyber-focused variant directly to developers back in September. Wider availability is expected to follow for paid API customers and Google AI Ultra subscribers first, then broader developer and enterprise access after that.
The timing lines up with OpenAI’s own push into this territory. Google is framing Argon’s CWE-bench score as a direct answer to GPT-6 Astra’s cybersecurity benchmark claims from earlier this month, and the gated rollout suggests both companies are treating autonomous vulnerability-finding capability as something that needs a slower hand than a typical model launch.
Pricing Once It Opens Up
Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. Google says that climbs to $4 and $20 respectively once the introductory period ends, so teams planning around Argon for anything beyond a pilot should budget for the higher number, not the launch price.
What This Means If You’re Not a Cyber Defender Yet
Most businesses won’t get hands-on access to Argon for a while, and that’s fine. The more useful signal here is direction: Google is explicitly building for agents that can sit inside a codebase or a compliance workflow for hours at a time, not just answer one question well. If you’re already running AI into client-facing or internal tooling, it’s worth tracking how Argon’s eventual general release pricing compares to what you’re paying now before locking into a model for a long-running integration. For teams weighing whether a given AI feature is worth building versus waiting for the next model generation to make it cheaper, that’s exactly the kind of tradeoff our AI development work runs into regularly.
