Model release · October 3, 2026
Google DeepMind has launched Gemini 4 Argon, its new top-tier model, with access starting through the Fairwind Program for vetted cyber defenders and through the U.S. government’s voluntary pre-release review. A wider opening to AI Ultra subscribers and API tiers is planned after that gate.
Briefings describe a 1-million-token output window and pricing of $2 per million input tokens and $10 per million output tokens. Google has cited 77.9 percent on DeepSWE v1.1 and hardened sandboxes aimed at prompt injection. A company spokesman has called it the most performant model yet for complex workloads, positioned against frontier coding and cyber benchmarks.
The restricted start is deliberate. Argon is built for software engineering, cybersecurity defense, and professional workflows, and Google scrapped a delayed Gemini 3.5 Pro along the way. Early access is going to defenders first, while the company participates in the government’s pre-release process.
That sequencing matters more than the benchmark line. A model sold for cyber work is also a model that can find weaknesses. Google is treating the first users as part of the safety case, not only as a marketing cohort.
Key takeaway: Gemini 4 Argon is live first for vetted cyber defenders, with a 1-million-token output window and $2/$10 pricing, before a broader subscriber and API rollout.
Photo: Mario Gogh via Unsplash
Sources
- Google DeepMind announcement coverage, late September–October 1, 2026
- AI Weekly and AI Briefing summaries, October 1–3, 2026
