Gemini 4 Argon released: 1 million token output, initially only for trusted security teams.

动察 Beating AI Brief: Google DeepMind releases its new-generation flagship model Gemini 4 Argon. It focuses on long-horizon complex tasks, covering software engineering, finance and legal work, as well as cybersecurity. The model is not yet open to ordinary users; the first batch is provided only to trusted cybersecurity teams through the Fairwind Program. It will later be gradually opened to paid API customers and Google AI Ultra users.

One of Argon's biggest changes is output length. The single-output limit has been raised from the previous 64K tokens to 1 million tokens, allowing the model to continue reasoning and execute longer steps within a single task. Google says it has already been used internally for large-scale code migration, algorithm optimization, and data center memory optimization.

In official evaluations, Argon scored 77.9% on DeepSWE v1.1, higher than GPT-6 Astra's 74.1% and Claude Opus 5.5's 74.2%; it also ranked first on Vals Index and AutomationBench. But it does not lead across the board, and on FrontierSWE, Terminal-Bench 4.0, and some scientific tasks, it still trails GPT-6 Astra or Claude Opus 5.5.

Google also placed heavy emphasis on strengthening cybersecurity capabilities. Argon can automatically discover, verify, and patch software vulnerabilities, and the version provided to trusted defense teams will even remove some cybersecurity restrictions to unleash its full capabilities. Before the official release, Google is still continuing to test risks such as abuse, prompt injection, and Agent boundary violations. At the initial stage of the API launch, pricing is $2 per million input tokens and $10 for output, after which the price doubles to $4 and $20.