Google launches Gemini 4 Argon for trusted cyber defenders

Google has announced Gemini 4 Argon, a model it is rolling out first to a set of trusted cyber defenders through its Fairwind Program rather than to the general public. Pricing starts at $2 per million input tokens and $10 per million output tokens, with cached inputs at a 95% discount, rising to $4 and $20 after the introductory window.

The headline capability is a 1 million token output limit, up from 64,000 on the previous generation, which Google says lets the model reason through long, complex problems in a single continuous pass. Google engineers have already used Argon internally to improve a quantum computing baseline by 40% and to find memory optimisations that freed more than 300 TiB across its data centres, with an estimated 500 TiB to 1 PiB of further savings available.

Argon is also being used to rewrite C and C++ code in Rust, including Google’s Fuchsia Zircon kernel, which has more than 800,000 lines of code. For the open-source libgav1 video decoder it replaced 32,000 lines of SIMD code, producing a version that runs 2.7 times faster than the previous Rust build with identical output and memory safety.

For trusted security teams and Google’s own engineers, Argon runs without its usual cyber safety restrictions so it can hunt for vulnerabilities. Wiz is already using it through its Scan for Good program, where Argon found a critical flaw in healthcare software used by hospitals worldwide that earlier AI models had missed. It scored 68% on CWE-bench v1, tying for first place.

Google says it is hardening defences on four fronts before any broader rollout: blocking misuse for cyber or CBRN attacks while allowing legitimate dual-use research, improving resistance to indirect prompt injection, and sealing off sandboxed environments used for high-risk training and testing. Argon’s chain-of-thought and actions are monitored in real time, with execution halted if the model drifts beyond the user’s request, and Google says it does not feed what that monitoring catches back into training.