The model that arrives as a clearance

Koray Kavukcuoglu announced Gemini 4 Argon as Google's new frontier model for long-horizon coding, enterprise knowledge work such as legal and finance, creative writing, and cybersecurity defense. Outside Google, the first external lane is the Fairwind Program, which opened on 3 September around Gemini 3.8 Flash Cyber and which reporting puts at more than 650 organizations, including CrowdStrike and Palo Alto Networks. Fairwind members and Google's internal teams get a version of Argon without cyber guardrails so they can use frontier defensive capability to find, validate, and patch vulnerabilities. Google cites early Wiz Scan for Good work and a 68% score on CWE-bench v1, tied with GPT-6 Astra on Google's own comparison. Output token limit rises to 1 million, up from 64 thousand. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off; after the intro period those rates become $4 and $20. Google says broader release starts with paid API customers and Google AI Ultra subscribers, with no public date. The announcement landed a day after OpenAI's DevDay launches.

1. 01On 30 September 2026 Google announced Gemini 4 Argon, rolling out first to trusted cyber defenders through the Fairwind Program, while taking a phased approach and engaging the U.S. government's voluntary process for pre-release model access.Google DeepMind blog, Koray Kavukcuoglu, 30 Sep 2026
2. 02Google describes Argon as frontier on complex workflows across real-world software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense, with creative writing also named for the broader product.Google DeepMind blog, 30 Sep 2026
3. 03Fairwind opened 3 September with Gemini 3.8 Flash Cyber. Reporting puts membership above 650 organizations, including CrowdStrike and Palo Alto Networks. Fairwind and Google internal teams receive Argon without cyber guardrails for defensive use.SiliconANGLE, 30 Sep 2026; Google DeepMind blog, 30 Sep 2026
4. 04Google says Argon can autonomously find, validate, and patch critical software vulnerabilities. On CWE-bench v1 Google reports a 68% remediations score, tied with GPT-6 Astra. Wiz is cited using Argon via Scan for Good, including discovery of a critical exposure in healthcare software that prior frontier models missed, per Google.Google DeepMind blog, 30 Sep 2026; TechCrunch, 30 Sep 2026
5. 05Output token limit expands to 1 million tokens, up from 64 thousand. Google reports state-of-the-art on DeepSWE v1.1 at 77.9%, leading position on the Vals Index, and 91.7% on LVBench for long video understanding.Google DeepMind blog, 30 Sep 2026
6. 06Introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off input price. After the introductory period, $4 per million input and $20 per million output apply. Broader availability is described as starting with paid API customers and Google AI Ultra subscribers; no public date was given.Google DeepMind blog, 30 Sep 2026
7. 07The Argon announcement came one day after OpenAI's DevDay, where OpenAI launched products including Dots and GPT-6.1 Sol.The Verge, 30 Sep 2026

We assess, with high confidence, that the headline for this desk is not that Google shipped Gemini 4 to the market. It is that Google shipped a frontier class it will not yet sell broadly. Phased access, Fairwind, and the U.S. voluntary pre-release process are the product for this week. The benchmarks and the price card are the brochure for a later SKU.

Treat Fairwind as a different object from the public product. Defenders in that cohort get cyber guardrails removed so the model can do offensive-capable defensive work under Google's trust boundary. Everyone else inherits a wait. Procurement that rewrites Q4 roadmaps around Argon API access this week is buying a press release.

The timing next to OpenAI DevDay matters for narrative, not for your contract. Until Google names a cohort and a date, Argon is announced capacity, not available capacity.

• Access is the scarce resource this week, not model weights on a slide.
• Fairwind membership is the live lane for Argon's cyber capabilities; the paid API is still a promise.
• Budget intro $2/$10 separately from the later $4/$20 so finance does not treat the discount as permanent.
• Independent evals will matter more than Google's chart once outsiders can run the model.
• Rival launches this week change the press cycle; they do not change who can call Argon today.

If a capacity plan, vendor shortlist, or Q4 roadmap assumed Gemini 4 API or Google AI Ultra access after this announcement, mark Argon as announced-not-available until Google names a cohort and a calendar date. File Fairwind as the only external lane that is live. Put the introductory $2/$10 rates next to the later $4/$20 rates in any cost model so the discount does not silently become the plan.

• Expansion of Fairwind or the first named paid API / AI Ultra availability date.
• When Google ends introductory pricing and moves to $4/$20.
• Independent (non-Google) Argon evaluations on coding, agents, and cyber defense.
• Any public note from the U.S. voluntary pre-release process that clarifies timing.
• Whether Google publishes a concrete broad-availability window rather than 'as soon as possible.'

• Google DeepMind blog - Koray Kavukcuoglu, 30 Sep 2026: Introducing Gemini 4 Argon
• TechCrunch - Lucas Ropek, 30 Sep 2026: Argon release, cyber focus, Fairwind
• The Verge - 30 Sep 2026: trusted cyber defenders first; DevDay timing
• SiliconANGLE - Duncan Riley, 30 Sep 2026: Fairwind scale, pricing vs Anthropic, CWE-bench