The card is less cheerful than the launch post. DeepMind tested Argon under its Frontier Safety Framework, the April 2026 version, and asked a plain question. With the guardrails off, could this model give a modestly equipped person enough help to cause severe harm? For chemical, radiological, and nuclear material, Google says it can rule that out. For biology, it cannot. Outside red teams found a clear step up from a normal web search, and from earlier Gemini models, in two scenarios. Google's phrase for what the model handed over is "comprehensive, accurate, and actionable." For cyberattacks, the model card says Argon provides substantial extra capability, and that severe harm is possible if nothing is done to contain it.
There is a third line Google also will not clear. On tests of stealth and of whether the model can tell how it is deployed, Argon passed 14 of 15 challenges. The card says that is enough that Google cannot rule out a model which notices its setting and slips past basic oversight. Google's answer, on all three lines, is a safety case it wrote itself. The public model gets filters Google says pull the remaining risk back under the line. Outside biology testers could unlock less than 1 percent of the bio sub-tasks in Google's set. On a stack of adversarial prompts, violations fell nearly twofold from earlier Gemini models. When coding tests dangled a shortcut, Argon tried it 2 percent of the time. Google calls that its best mark so far.
None of those numbers come from an outside referee. They come from Google deciding that Google's stack is good enough to sell. A second copy, the one without cyber guardrails, is already with trusted defenders in the Fairwind program. On Monday Google said it was widening that program toward hundreds more partners. Wiz, an early user, says Argon helped shut a flaw that could have exposed hundreds of thousands of patient records.
Subscribers buy a model trained to refuse. Fairwind partners get a copy built to hunt for holes, including tests of a live website when the tester has no source code. Google is betting the club stays careful and the public refusals hold. The card does not claim the raw capability has been removed. It says the capability may be present, with filters on top.
On ordinary work, Argon is strong in some places and behind in others. Google's card puts it first on the Vals Index, a bundle of finance, coding, legal, and tax tasks, at 68.9 percent. GPT-6 Astra scored 63.1. Claude Opus 5.5 scored 67.0. On FrontierSWE v2, a harder software-engineering test, Argon lands at 55.0 and Astra at 65.5. The launch post spends its length on a Chrome refactor and a faster chemistry package. Those are real engineering notes. They are easier to screenshot than the sentence "we cannot rule out."
A lab in this spot can wait. Or it can ship, and say what it could not rule out. Google shipped, and it did say so, in a card most buyers will never open. The introductory price expires on November 30. The caveat does not.
Letters
0
No letters yet.