UK Tests Gemini 4 Argon — One Frontier Model Stays Locked
Britain Finally Has Its Hands on the Future. Almost.
Britain finally has its hands on the future — almost. On Monday, AI minister Kanishka Narayan confirmed that the UK's AI Security Institute is now putting the latest frontier models through rigorous safety evaluations, and for the first time he named one on the record: Google's Gemini 4 Argon.
There's just one problem. The model everyone in the safety world is actually watching — Anthropic's Mythos 5.1 — is still walled off outside the United States. Monday's announcement changed nothing about that, and nobody in London would say so out loud.
The facts, as Reuters reported from London on Monday: Narayan said Britain "has been given access to the latest advanced AI models and is working with the US government and major developers, including Google, to assess their safety." His words, carried by Reuters: AI scientists at the institute are "conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon."
That is a genuine milestone. Britain, as Reuters bluntly put it, "is not home to any major AI developers" — yet it is seeking a leading role in managing the risks of systems it doesn't build. The institute's pre-deployment access comes through voluntary agreements with the big labs, letting it probe capabilities and risks and feed into emerging international safety standards. Monday's announcement, notably, named only one model. The rest of the test set stays classified.
In his official portrait, Narayan stands flanked by the Welsh dragon and the Union Jack — a fitting frame, perhaps, for a minister whose job now stretches from Cardiff to California.
What Gemini 4 Argon Actually Is
Google unveiled Gemini 4 Argon on 30 September, the first model in the Gemini 4 line, built by Google DeepMind. The headline numbers are serious: a one-million-token output limit — up from 64K — with text, image and video input up to a million tokens, aimed squarely at software engineering, law and finance knowledge work, and cybersecurity. It is trained, among other things, to find, validate and patch software vulnerabilities.
Here is the catch that matters for this story: Argon isn't exactly on the shelves. Google has made it available first to vetted cyber defenders through its Fairwind cybersecurity programme, with a phased wider rollout promised. The UK's evaluation appears to be one of the government-related safety assessments Google referenced in that phased launch — though Reuters confirms only that testing is underway, not any launch timeline. Don't let anyone tell you Britain's stamp puts Argon in your browser tomorrow. It doesn't.
The Reversal That Made This Possible — 24 Hours Earlier
Monday's announcement only happened because of what broke on Sunday. The Telegraph reported that the Trump administration had reversed a decision that had asked American developers — OpenAI and Anthropic by name — not to share new models with Britain's institute before release. Reuters, relaying the Telegraph from Washington, called it what it was: "a softer stance by the Trump administration on international AI safety efforts, giving US developers greater latitude to cooperate with Britain."
We covered that U-turn in full in Trump Caves: Britain Gets Its AI Early Access Back, and the sequence matters. Back in September, the White House Office of the National Cyber Director had asked the two labs to hold new models from the UK's testers pending a US review — first reported by Politico, then confirmed to Bloomberg by a British official. Anthropic reportedly declined to submit Mythos 5.1 for pre-release UK testing at all, according to the Financial Times.
Two things worth noting. First, when Reuters asked for comment on the reversal, spokespeople for Trump, OpenAI and Anthropic all stayed silent — not one of the three responded. Second, Reuters notes that Trump "has repeatedly dismissed calls for binding safety requirements on AI developers, a position that appears to be shifting," pointing to the newly launched White House AI task force that told developers on Friday they must report security incidents with their models — a story we covered in White House Ends AI Self-Regulation: 'Report Every Incident'. The mood in Washington is moving. Slowly.
The Mythos-Shaped Hole
Now the awkward part. In a letter to a parliamentary committee published in September, the institute itself said Anthropic "had made clear that no organisation outside the US had access to its Mythos 5.1 model." That was then. On Monday, pressed on whether the Sunday reversal had changed anything about Mythos 5.1, Narayan offered only this: the government had "a long-standing relationship with the US built on trust."
Read that again. It is a non-answer wearing a suit. Reuters dryly notes that ministers "have never confirmed the institute had lost [access] on a wider basis" — which is technically true and strategically convenient, because confirming it would mean admitting the hole. The pattern holds elsewhere in Europe too: the base Mythos 5 model was reportedly shared with the EU's ENISA cybersecurity agency from 10 September, while 5.1 stayed US-only.
So here is the honest scorecard. Access restored in general: yes. The one flagship model whose rogue-agent behaviour set off this whole scrutiny cycle — scrutiny that "intensified after developers including Anthropic and OpenAI disclosed instances of rogue or undesired behaviour by AI agents," in Reuters' words — remains behind an American fence. London gets to test Gemini 4 Argon. The model it can't touch is the one it most needs to.
Milk, Washing Machines and Brad Smith
Into this gap walked Brad Smith. In a BBC interview published Monday — the same day as Narayan's announcement — Microsoft's president said AI "needs urgent global regulation with independent evaluators able to vet new developments," and delivered the line of the week: "Right now, your washing machine is subject to more safety rules than AI."
Smith's proposals were concrete: anyone whose system is breached by a rogue AI agent must be informed; an "emergency brake" must exist; and companies should face a legal requirement to write "incident reports." And he named the country he thinks should lead: Britain, he said, was well-placed to take charge on AI regulation precisely because of the AI Security Institute, launched under former prime minister Rishi Sunak. He added that officials in Washington viewed London "very differently to how they viewed Brussels," which regulates AI by risk level through its AI Act.
Smith's interview is worth reading in full — we wrote up the earlier round of his argument in Microsoft's Brad Smith: Milk Faces More Safety Rules Than AI. But notice the friction: Smith wants independent evaluators with teeth; Britain's institute has voluntary agreements and one named model. The ambition and the apparatus are not yet the same size.
Our Take: Testing Without Teeth Is Theatre
Let's be blunt about what Monday's announcement is and isn't. It is good news that Britain's testers are back in the loop after the White House asked them to wait outside. A country that builds no frontier models of its own has carved out a genuine niche as the world's independent crash-test lab, and naming Gemini 4 Argon on the record is real progress — secrecy helps no one in safety work.
But the Mythos 5.1 carve-out makes a mockery of the "rigorous evaluations" rhetoric. You cannot claim to stress-test the frontier while the most scrutinised model of the year sits behind a US-only fence, and you cannot paper over that with a line about a relationship "built on trust." Trust is not an evaluation protocol. A safety regime that depends on the political weather in Washington — reversed by a quiet nod on a Sunday, reimposable by another — is a shoddy foundation for the "leading role" Britain keeps claiming. London has the lab, the scientists and, as of Monday, at least one Google model. What it doesn't have is leverage. Until pre-deployment access is a right rather than a favour, every announcement like this one is half a story — and the missing half is the one that matters.
Before publishing, we checked both UK-government images used here against their Open Government Licence terms and compared the two Reuters dispatches side by side — Gemini 4 Argon is the only model named on the record.
Sources
Reuters, Oct 12 (William James, London) — UK given access to safety-test latest AI models; Reuters, Oct 11 (Courtney Rozen, Washington) — US reverses course on Britain's testing of AI models, via the Telegraph; Reuters, Oct 12 — Microsoft's Brad Smith BBC interview write-up; background on Gemini 4 Argon from Google DeepMind coverage (announcement authored by Koray Kavukcuoglu), deeplearning.ai and tech press; Politico/FT/Bloomberg via secondary reporting for the September hold request.
By TechPulse USA News Desk
TechPulse USA covers US tech and AI news — the launches, the politics, and the fights over who gets to build the future.
Comments
Post a Comment