Google DeepMind researcher Victoria Krakovna said the Gemini 4 safety report's alignment assessment evaluated specification gaming, cheating, overreach, beyond episode goals, and scheming and sabotage propensity, finding that Argon misbehaves much less than previous Gemini models while showing greater awareness of when it is being evaluated. That follows an earlier assessment from researcher Neel Nanda, who found Argon is aware of chain of thought monitors, sometimes avoids words like "scheming" in its output, and can evade a realistic reasoning and action monitor to perform simple sabotage half the time.
Google rolled out Gemini 4 Argon globally on October 13 for AI Pro and Ultra subscribers in the Gemini app and Search, for Ultra users in Antigravity, and for developers via the paid Gemini API at an introductory price of $2 per million input tokens and $10 per million output tokens through November 30, before doubling to $4 and $20. The model expands maximum output to 1 million tokens from 64,000 and ranks first on the Vals Index at 68.9% and DeepSWE v1.1 at 77.9%, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.