The launch

Google announced Gemini 4 Argon on September 30, its first new in-house flagship in over seven months. The model is built for long, multi-step work: coding, financial and legal research, and cybersecurity defense. It can generate up to a million output tokens in a single response, up from 64,000 in the previous generation.

Access, limited to defenders

The part that matters more than the benchmarks. Argon is currently available only to trusted cyber defenders through Google's Fairwind Program, which launched in early September with more than 650 partners. Google says these partners and its own internal teams get the model without cyber guardrails, so it can find, validate, and patch vulnerabilities at full capability. Paid API customers and Google AI Ultra subscribers are next in line. Google has given no date.

The benchmark table, read carefully

Google's launch materials include a 19-row benchmark comparison against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5. Argon comes out on top in 13 rows, ties one, and loses five. It finishes last of the four on FrontierSWE v2 and Terminal-bench 4.0, two coding benchmarks developers actually watch. The independent number is less flattering. Artificial Analysis scores Argon (high) at 53 on its Intelligence Index, tied with Astra and Fable. Claude Opus 5.5 leads at 58. Argon's real advantage is price. It costs roughly a third as much per task. At launch it is priced at $2 per million input tokens and $10 per million output tokens, with cached input 95 percent off, matching the list price of OpenAI's GPT-6.1 Sol.

What this actually means

For anyone outside the Fairwind Program, Argon is a press release, not a product. That may change, but Google's own timeline offers nothing to plan around. The interesting decision is Google's. A flagship restricted to defenders first suggests the company considers the cyber capability its most valuable and its most sensitive part of this model.

Sources

  1. [1] securityweek.comRead source
  2. [2] theubj.comRead source
  3. [3] dev.toRead source