Google DeepMind launches Gemini 4 Argon with 1M-token output for cybersecurity preview

Google DeepMind introduced Gemini 4 Argon, a frontier model positioned for complex workflows in coding, enterprise knowledge work, and cybersecurity defense. The model is currently accessible only to government users and trusted cyber defenders within the Fairwind Program, with broader access promised “as soon as possible.”

Argon introduces an industry-first 1M-token output limit via a feature called Long Decode Continuation, which pauses long responses and resumes them across API calls. This capability is designed to support extended agentic workflows and long-form outputs.

Google DeepMind reports that Argon achieves state-of-the-art results in 13 of 19 published benchmarks. On DeepSWE, it scores 77.9%, compared to 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. It also ranks first on AutomationBench-AA at 77.5% and scores 57.6% on Terminal Bench 4.

The model’s pricing is set at $4 per 1M input tokens and $20 per 1M output tokens, with a 50% introductory discount bringing costs to $2/$10. Cached input tokens receive a 95% discount. Despite these savings, Argon’s average output per task is 62K tokens, higher than Astra’s 27K, which affects total task costs.

Internal deployments of Argon-based agents reportedly freed more than 300 TiB of data-center memory and are migrating over 800K lines of C/C++ kernel code to Rust. Agents also replaced 32K lines of SIMD code with safe Rust, improving the existing Rust port’s speed by 2.7x while maintaining identical output.

Research use cases include internal agent loops built on Argon that helped complete the CK conjecture. Argon scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra and edging out GPT-6.1 Sol (52).

On agentic benchmarks, Argon ranks eighth overall and first for steerability in a preliminary evaluation of 3,000 sessions. It also leads on Text Arena at 1525 and ranks eighth in Code Arena WebDev at 1679. However, it trails on some specialized evaluations, such as Harvey’s legal benchmark, where it scores 19.6% compared to Muse Spark 1.2’s 25.42%.

Observers have raised questions about the published numbers, including potential preference-data benchmaxxing and critiques of specific benchmark figures like DeepSWE. Despite skepticism, Argon’s performance and features underscore Google DeepMind’s emphasis on long-context agentic capabilities and cybersecurity applications.