Google limits Gemini 4 Argon to vetted cyber defenders

Google DeepMind introduced Gemini 4 Argon as its first Gemini 4 model, with access initially limited to select trusted cyber defenders through the Fairwind Program and Google’s internal teams. The model is positioned for sustained reasoning across long workflows, including software engineering, corporate knowledge work, finance, legal tasks and cybersecurity, with a maximum of 1 million output tokens.

Argon leads many of Google’s published comparisons, ranking first alone in 12 out of 18 benchmarks and tied for first in one. Independent results are more nuanced: Vals AI ranked it 1st out of 41 models on the Vals Index with 68.90%, while Artificial Analysis placed it in the top group with an Intelligence Index score of 53. The model appears especially strong in long context, knowledge work, finance, legal analysis, automation, multimodal work and cyber defense.

Weaknesses remain in terminal-based coding, some software engineering tests, PC operation and certain science-agent tasks, where GPT-6 Astra, Claude Opus 5.5 or Claude Sonnet 5.5 outperform it. Google lists introductory pricing at $2 per 1 million input tokens and $10 per 1 million output tokens, rising after the introductory period to $4 for input and $20 for output.