Google Unveils Gemini 4 Argon, Limits Access to Cybersecurity Experts

Google announced Gemini 4 Argon on September 30, billing it as the company's first true flagship AI model in nearly a year. According to Google's self-published benchmarks, Argon outperforms Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra across most categories, particularly in software engineering, knowledge work, and cybersecurity defense. Independent verification of these results is not yet available.

The model's standout technical feature is a one-million-token output limit, a significant increase from the previous 64,000 tokens. Google states this allows Argon to solve complex, multi-step problems in a single pass. The company reports internal use by thousands of employees for coding and research, citing memory optimizations freeing over 300 tebibytes of RAM and a video decoder running 2.7 times faster than its predecessor.

“Argon is fundamentally changing the way we work and build at Google.”

Access is currently restricted to selected cybersecurity experts within Google's 'Fairwind' program, launched September 2 for governments and trusted partners. Google says this controlled release allows time to adjust safety filters before a wider public launch, which is promised "as soon as possible." The U.S. government is also reviewing the model under a voluntary commitment process.

Pricing for the introductory launch is set at $2 per million input tokens and $10 per million output tokens, which Google notes is half the cost of Anthropic's Claude Opus 5.5. Google's benchmark data shows Argon leading on the Vals Index with a 68.9% score and reaching 77.9% on DeepSWE v1.1 for coding tasks.