Google launches Gemini 4 Argon: 1M-token output, first to trusted cyber defenders

Google DeepMind launched Gemini 4 Argon on September 30, its first flagship model in nearly 10 months since Gemini 3 in November 2025. DeepMind Chief AI Architect Koray Kavukcuoglu described Argon in an official blog post as the next stage of frontier intelligence.

Argon targets long-horizon software engineering, enterprise knowledge work, and cyber defense. On Google's published benchmarks, Argon posts 77.9% on DeepSWE v1.1, beating Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%), ties for first at 68% on CWE-bench v1 (vulnerability remediation) with Gemini 3.8 Flash Cyber, lands at 51.3% on AutomationBench for end-to-end business workflows, and 84.2% on GraphWalks 256k-1M. The model lifts the single-shot output limit to roughly 1 million tokens, up from 64,000 in prior generations, to handle large codebases, long documents, and multi-stage tasks.

Introductory API pricing is $2 per million input tokens and $10 per million output (with cached input at $0.10), rising after the intro window to $4 input and $20 output, matching the list price for Claude Opus 5.5. Bloomberg reported on September 30 that some Google employees privately feel Argon performs less well on real coding tasks than the benchmark scores suggest. Google pushed back on the characterization, saying the description is inaccurate and there is large consensus inside the company that the model is at the frontier.

Access rolls out in phases: trusted cyber defenders in the Fairwind program (including security firm Wiz through its Scan for Good initiative) come first, followed by paid API customers and Google AI Ultra subscribers, alongside participation in the US government's voluntary pre-release model evaluation process. Google did not announce a specific timeline for opening access to developers and consumers. The 1M-token output cap and Fairwind-first rollout are read as Google's latest attempt to bind capability release to safety gating.