Google has launched Gemini 4 Argon, the first model in its new Gemini 4 series, with initial access limited to trusted cybersecurity defenders and internal teams. The release was announced on September 30 through Google's Fairwind Programme, which provides select government partners and organisations with early access to advanced AI tools for cyber defence. Argon is designed to autonomously detect, verify and patch critical software vulnerabilities, and it can generate up to 1 million tokens in a single output—far exceeding the 64,000-token limit of earlier models. This expanded capacity allows it to handle large-scale software engineering tasks, such as migrating extensive codebases from C and C++ to Rust, including work on the Zircon kernel in Google's Fuchsia operating system.
Google reports that Argon scored 77.9% on the DeepSWE v1.1 benchmark for software engineering, outperforming Anthropic's Claude Opus 5.5 at 74.2% and OpenAI's GPT-6 Astra at 74.1%. On CWE-bench v1, which tests vulnerability detection and repair, Argon achieved 68%, matching GPT-6 Astra. These results come from Google's internal evaluations, not independent testing. The company is using Argon internally to analyse data-centre performance, where it has identified memory optimisations expected to free over 300 tebibytes of memory, with total savings projected between 500 tebibytes and 1 pebibyte. Quantum computing researchers have also used Argon to improve algorithm efficiency, surpassing a published baseline by 40% in minutes. Access for paid API users and Google AI Ultra subscribers will come later, but no rollout date has been set.
The decision to restrict access to Gemini 4 Argon despite its high benchmark scores suggests Google recognises the dual-use risk of an AI that can independently patch or exploit software flaws. Performance claims based solely on internal tests remain unverified outside its ecosystem until broader deployment allows independent validation. When wider access begins, real-world performance will determine whether Argon's capabilities match Google's reported outcomes.
Editorial note: AI-assisted opinion, not established fact. Full disclaimer →
