Google launched Gemini 4 Argon on September 30, 2026, as the first model in its new series. The AI is designed to autonomously identify, validate, and patch software vulnerabilities. Beyond cybersecurity, the model supports complex workflows in software engineering, law, and finance.
The model features an output capacity of 1 million tokens per response, a significant increase from the 64K limit of previous versions. In testing, Argon achieved a 77.9% score on the DeepSWE v1.1 benchmark, outperforming OpenAI GPT-6 Astra at 74.1% and Anthropic Claude Opus 5.5 at 74.2%.
Internal applications include migrating over 800,000 lines of code in the Zircon kernel to Rust, which saved 300 TiB of memory in Google data centers. The model also scored 68% on the CWE-bench v1 vulnerability remediation test.
Access is currently restricted to the Google Fairwind Program, which includes over 650 vetted organizations such as CrowdStrike, Palo Alto Networks, and Wiz. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens, with plans to increase rates to $4 and $20.
