Overview
- Google has begun a controlled rollout of Gemini 4 Argon to trusted cybersecurity partners through its Fairwind program, with wider paid access planned for Google AI Ultra subscribers and API customers.
- Argon introduces an unprecedented 1,000,000-token output limit that lets the model sustain very long, multi-step reasoning in a single run for tasks such as large-scale code migration and long-form document or video analysis.
- Google reports benchmark-leading scores in programming and remediation tests, including 77.9% on DeepSWE v1.1, 91.9% on Vibe Code Bench, and a tied top result of 68% on CWE-bench v1 for vulnerability remediation.
- In pilot tests with security firm Wiz, Argon reportedly found a previously undetected critical vulnerability in global hospital software and is trained to locate, validate and autonomously repair flaws across more than 20 programming languages.
- Google says Argon runs with built-in prompt-injection protections, real-time monitoring, and staged access controls while internal deployments have already sped automated workflows by about 40% over published baselines.