Overview
- Gemini 4 Argon, which Google announced Wednesday, Sept. 30, 2026, is in a controlled rollout to vetted cybersecurity partners through the Fairwind Program and to Google’s internal teams.
- Google’s published benchmarks show Argon leading on many coding, knowledge‑work and cyber tests compared with OpenAI and Anthropic models, though it trails on some coding benchmarks.
- Argon expands single‑run capacity to a 1 million output‑token limit, letting the model pursue much longer, single‑trajectory reasoning and generation than prior Gemini releases.
- Google says Argon is trained for defensive cyber work and is already used by Wiz in Scan for Good, which Google reports found a critical healthcare‑software vulnerability, and internally to free more than 300 TiB of memory and help large code migrations.
- The model will be offered first to paid API customers and Google AI Ultra subscribers after the safety testing phase, with introductory API pricing set at $2 per million input tokens and $10 per million output tokens and independent verification of claims still pending.