Quick answer

Google DeepMind announced Gemini 4 Argon on September 30, 2026. It is a frontier model built for long-horizon work: real-world software engineering, enterprise legal and finance tasks, and cybersecurity defence. Context is 1M tokens and, unusually, so is the output limit, up from 64K, so one run can keep thinking and writing for hundreds of thousands of tokens. Introductory API pricing is $2 per million input tokens and $10 per million output, rising to $4/$20 after an unannounced promotional period, with cached input at $0.10. At launch it is available only to vetted defenders through Google's Fairwind Program; paid API customers and AI Ultra subscribers are next, with no date given. Model ID: gemini-4-argon.

Three frontier-tier launches in nine days, all at the same $2/$10 introductory price: Anthropic's Sonnet 5.5 on September 28, OpenAI's GPT-6.1 Sol on September 29, and now Google's Gemini 4 Argon. Argon is the one you cannot use yet, and also the one aimed at the longest jobs.

What makes Argon different

  • Output limit of 1M tokens: most frontier models cap a single response at tens of thousands of tokens; Argon can run a coding or research task for far longer without stopping and restarting
  • Long Decode Continuation: a mechanism to pause and resume an extended response, which matters when a run is interrupted
  • Multimodal input: text, image, video, and speech in; text out
  • Stated targets: software engineering, legal and finance knowledge work, and cyber defence, rather than general chat

The benchmarks Google reported

Google says Argon leads on 14 of 19 benchmarks it tested, citing figures such as 19.6% on Harvey's Legal Agent benchmark, 65.4% on Vals Finance Agent v2, 77.9% on DeepSWE v1.1, 84.2% on GraphWalks at 256K to 1M tokens, and a tie at 68% on CWE-bench v1 for cybersecurity. Every one of these is Google-reported, several are on benchmarks Google partners maintain, and independent evaluations were not available at publication. Treat the direction as meaningful and the numbers as provisional.

Who can use it, and when

Access is through the Fairwind Program, which Google describes as serving high-priority defenders: governments, healthcare, telecommunications, and critical-infrastructure providers, with more than 650 partners. General developers are excluded for now. The public Gemini API returned a not-found error for the model ID on October 1. Google says paid API customers and Google AI Ultra subscribers come next, then broader access, without dates. This mirrors OpenAI's staged rollout of GPT-6 Astra in September, and reflects the cybersecurity capability both companies now treat as a gating factor.

Pricing in context

  • Introductory: $2 input / $10 output per million tokens, cached input $0.10 (a 95% discount)
  • Standard after the promotion: $4 / $20, with no end date announced
  • Same introductory tier as Sonnet 5.5 and GPT-6.1 Sol; half the price of GPT-6 Astra
  • A 1M-token output at $10 per million is a $10 single response — budget accordingly for long agent runs

Related reading

GPT-6.1 Sol vs Sonnet 5.5 vs Gemini 4 Argon Claude Sonnet 5.5 released Gemini 3.8 Flash launch

If you build on Google Cloud and your agents run for hours, Argon is the model to plan for. If you need something today, Sonnet 5.5 and GPT-6.1 Sol are available at the same price.

Bottom line

Gemini 4 Argon is a bet that the next frontier is endurance, not just intelligence: models that can work for a very long time in one run. The pricing matches its rivals; the access does not yet. Watch for the API opening before you re-architect anything around it.