Google Unveils Most Powerful Model Yet, Lets Approximately Nobody Use It
Gemini 4 Argon leads 13 of 18 disclosed benchmarks, and is available to a carefully vetted guest list — which is one way to win a leaderboard.
Google has released the most powerful AI model it has ever built. You cannot have it.
This is the central tension of Gemini 4 Argon, announced September 30 with the confidence of a company that has just topped the benchmark charts and the discretion of a company that would like you to stop asking to see it. It is a launch in the sense that a restaurant opening is a launch when the only diners are the health inspector and the owner’s relatives.
Argon, per Google, is “beginning rollout to a set of trusted cyber defenders” through the Fairwind Program, a limited-access club the company set up in early September for governments, cloud customers, and security partners. Broader availability is planned “as soon as possible,” a phrase that in corporate English occupies the long, shimmering stretch between “Tuesday” and “we’ll circle back.”
The numbers, for those who enjoy a leaderboard, are strong. Argon leads or ties on 13 of the 18 benchmarks Google disclosed, which is the sort of statistic that is true and also invites the question of how the other five were chosen for display. On DeepSWE it posts 77.9 percent against Claude Opus 5.5’s 74.2. On Harvey’s legal agent benchmark it scores 19.6 percent, against GPT-6 Astra’s 5.4 — a margin so wide it suggests that either Argon is very good at law or Astra was, at the time of testing, thinking about something else.
It does not win everything. Astra still leads on FrontierSWE v2, 65.5 to 55.0, and Anthropic’s model keeps an edge on terminal-agent workflows. Google, to its credit, listed these too. Nothing builds credibility like admitting you lost a category in a press release that is otherwise a victory lap.
Argon also supports a one-million-token output limit, up from 64,000, meaning it can now produce roughly the length of a long novel in a single response. It is unclear who asked for this. Presumably the same people who asked for the reply-all button.
Then there is the price, which is where the “price war” tag earns its keep. Introductory rates are $2 per million input tokens and $10 per million output, with a 95 percent discount on cached input. After the introduction, the rate goes to $4 and $20 — a pricing structure with the shape of a free sample at the grocery store, if the grocery store were also the only one allowed to sell the food.
You will notice that this introductory price is a deal on a product that, for most people, does not yet exist. It is a sale on the idea of a model. Please line up.
Why the velvet rope? The honest answer is in the cyber capabilities. Argon is good at finding and fixing software vulnerabilities — Google says it caught a critical healthcare software flaw that earlier frontier models missed — and a model that is good at finding vulnerabilities is, by the unfortunate symmetry of the field, also interesting to people who prefer to find them for other reasons. Hence “trusted cyber defenders,” a category that Google presumably vets with the diligence of a bouncer who has read the incident reports.
The model is also moving through the U.S. government’s pre-release review, which, as regular readers know, is voluntary. Reporting indicates the launch landed a day after Google and other labs joined a voluntary safety agreement covering pre-release evaluations. Voluntary, once again, in the way that attending the meeting you scheduled is voluntary.
Google says it is strengthening safeguards in four areas: misuse, prompt injection, monitoring for misalignment, and hardening sandboxed environments. That is a real list, and a long one, and the fourth item in particular has the faint but unmistakable sound of lessons learned.
What we have, then, is a new genre: the benchmark-victory-by-appointment. The model is the best in the world, by Google’s own disclosed numbers, and the world will receive it at a time to be announced, pending review, after a safety checklist, once the right people have used it first. The trusted cyber defenders get the first look. Everyone else gets a blog post and a pricing table.
It is, to be fair, a perfectly sensible way to release a powerful tool that can also break into things. It is also a way to hold the crown without having to answer the door. Somewhere in Mountain View, a product manager is refreshing a leaderboard and has never been so confident, or so unavailable for comment.
Sources: