Google launches Gemini 4 Argon for defenders first
Google DeepMind announced Gemini 4 Argon on 30 September 2026. Fairwind defenders get it first, without cyber guardrails. No public API date yet.
The frontier model, announced on 30 September 2026, starts in Fairwind. There is still no public date.
On 30 September 2026 Google DeepMind announced Gemini 4 Argon, a frontier model (the company’s most capable line) for long software jobs, enterprise knowledge work such as legal and finance, and cyber defense. The blog.google post, by Koray Kavukcuoglu, SVP of Google DeepMind and Google’s Chief AI Architect, and the @GoogleDeepMind post on X say the same thing: the model goes first to trusted defenders through the Fairwind Program. Reuters and Axios covered it the same day. Google sells the model.
If you wanted Argon in the paid API, Vertex, the Gemini app, or as a consumer, what changed that day was the order of the line. Vetted defenders get in now. Everyone else waits, with no date.
First: what Fairwind is
Fairwind is DeepMind’s early access program for internal cybersecurity, incident response, or pentest teams (authorized break-in tests). It is a closed list, with phishing-resistant MFA (a login a fake email cannot copy), usage tracking, and a ban on redistribution or resale. It is not a public beta. Dual-use tasks, according to the program page, are limited to defense or research.
A token here is a chunk of text the model reads or writes. A million output tokens is a very long reply. Google compares that cap with the previous 64,000.
Who gets Gemini 4 Argon now
According to Google, Argon goes to a subset of Fairwind partners, on its own or inside CodeMender, the company’s code-fixing tool. The company says there are more than 650 partners worldwide, across governments, healthcare, and telecom. It has not published the roster. That headcount is Google’s.
A wider release to developers, companies, and consumers would come “as soon as possible,” starting with the paid API and Google AI Ultra subscribers. RUMOR: there is still no date for the API, Vertex, the Gemini app, or consumers. The official wording is “as soon as possible” and “rolling out soon.” Right now it is for a defense team already in Fairwind that got an invite. It is not for anyone hoping to try the model in the app or the API today.
Price and the million-token output cap
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input, when a chunk already paid for is stored and reused. After the intro window, $4 and $20. RUMOR: Google did not say how long the discount lasts.
The output cap rises to 1 million tokens. Google calls that “industry-leading.” That is a vendor slogan, not a third-party measurement. RUMOR: the input window, how much the model can read at once, is not in the post.
What the numbers show
Google leads the slice it put on its own blog: DeepSWE v1.1 at 77.9%, Zapier’s AutomationBench at 51.3%, and LVBench at 91.7%. In DeepMind’s methodology, Argon’s DeepSWE and LVBench scores are “self computed”: the company ran the test.
The official product table is mixed. On the Vals Index, a third-party leaderboard, Argon scores 68.9%. On CWE-bench v1, another third-party leaderboard (Collinear AI), it ties at 68% pass@1 with Grok 4.7 and GPT-6 Astra; Claude Opus 5.5 is at 67%. Collinear sells training corpora. Vals is an eval company. Both have a commercial stake.
Where Argon does not lead, still according to Google’s table: FrontierSWE v2 at 55.0%, behind GPT-6 Astra (65.5%) and Claude Opus 5.5 (62.3%); Terminal-bench 4.0 at 57.4%, behind Opus 5.5 (66.4%). Argon’s Terminal-bench score is also “self computed.”
No cyber locks for the closed group
Google says it trained Argon to find, validate, and fix vulnerabilities on its own and that, “for trusted defenders and our own internal teams at Google, we’ll be releasing Argon without cyber guardrails.” A guardrail here is the lock that refuses a dangerous request. Without it, the model can help find and exploit bugs. For a wider public, Google says it will refuse harmful cyber and CBRN requests (chemical, biological, radiological, and nuclear weapons) under its Frontier Safety Framework. That is a product promise, not an independent audit. Google says the model is in the U.S. government’s voluntary pre-launch access process. Reuters reported that the same day.
What we still do not know
There is still no date for the paid API. Internal claims about data center memory and a bug in hospital software are Google’s, with no attached paper and, in the healthcare case, no CVE (a public bug ID). RUMOR: details of that incident. What changed on 30 September is the order of the line: vetted defenders first, with the cyber locks off, and everyone else still waiting.
Sources
Official
- Google Blog, Koray Kavukcuoglu, Gemini 4 Argon, 30 Sep 2026
- @GoogleDeepMind on X
- Google DeepMind, Fairwind Program
- Google DeepMind, Gemini page and evals table
- Google DeepMind, Gemini 4 Argon evals methodology
- Google DeepMind, Frontier Safety Framework
Third-party leaderboards
Press