Google begins limited rollout of Gemini 4 Argon

Google describes Gemini 4 Argon as a frontier model for coding, enterprise knowledge work and cyber defense.[1] The company says it is built to sustain deep reasoning across complex, long-horizon workflows.[1] Its initial release is not a general rollout: Google says it is first making Argon available to trusted cyber defenders through the Fairwind Program.[1] Google also says it is participating in the U.S. government’s voluntary process for pre-release model access as it expands access gradually.[1] These statements outline both the model’s intended work and a staged approach to distribution.

Google says it plans to make Argon available to developers, enterprises and consumers, starting with paid API customers and Google AI Ultra subscribers.[1] The company set introductory prices at $2 per million input tokens and $10 per million output tokens.[1] Cached input tokens are priced 95% below the input-token price.[1] Google has also raised the output-token limit to 1 million, from 64,000.[1] For teams evaluating the model, these figures provide initial cost and capacity parameters, but the announcement describes a phased rollout rather than access for all users at once.[1]

Google says thousands of its employees highlighted Argon’s strengths in specialized coding tasks, deeper research and writing quality.[1] Its account describes work spanning changes to core libraries with tens of thousands of lines and work involving more than 800,000 lines in the Fuchsia Zircon kernel.[1] Google also says an Argon-assisted rewrite made its libgav1 Rust video decoder 2.7 times faster than the existing Rust port while producing identical video output.[1] These are company-reported examples; the evidence provided does not establish that the same results will apply to other codebases or workflows.

Google says Argon can autonomously find, validate and patch critical software vulnerabilities, and can conduct black-box penetration tests using a running website’s public behavior without source-code access.[2][3] In Google’s benchmark charts, Argon tied with Grok 4.7 and GPT-6 Astra at 68% on CWE-Bench v1; Google also reported 85.8% on real-world vulnerability discovery and 70.9% on its Wiz penetration-test benchmark.[3] Google says the model had a 0.7% attack success rate on the Gray Swan indirect prompt-injection benchmark and that its safeguards monitor chain of thought and actions, stopping execution when necessary.[3][1] These figures and safeguards are Google’s reported results and descriptions.