The afternoon a government switched off an AI

Last week a government switched off the most capable AI models on the planet in an afternoon. On 12 June the US ordered Anthropic to cut access to Fable 5 and Mythos 5 for everyone outside America. The order was aimed at foreign nationals, not the world, but to comply the company disabled both for every customer, worldwide. A frontier model used by hundreds of millions, gone by teatime. Whatever side of the table you sit on, that should stop you.

The order was an export-control directive from the US Commerce Department, citing national security. It barred all foreign nationals, inside or outside the US, including Anthropic's own foreign-national staff, from using the two models. The trigger, as Anthropic understands it, was a claimed way of jailbreaking a safeguard that stops Fable 5 being used to find software vulnerabilities. Anthropic says the jailbreak is narrow, that other public models including OpenAI's GPT-5.5 can do the same, and that recalling a model used by hundreds of millions over it makes little sense. It complied anyway.

Here is what changed my mind. I had assumed Anthropic were simply better at hype than the rest. It turns out you do not issue a national-security export order against a marketing campaign. The US does not want anyone outside America near this model, and governments do not move like that over a clever launch. They move like that over capability they believe is strategically dangerous, especially in cyber.

The ban is the proof that the hype was not hype.

Call it a small earthquake, because it is a wake-up call on two sides of the table.

Two sides of the table

On one side: guardrails, jobs and capability, and whether any of us are ready. For years, export controls were about chips, the hardware. This is the first time the US has reached past the hardware to switch off access to the intelligence itself. That tells you governments now treat frontier AI as a strategic asset, close to a weapon. And the reason given, that a model is capable enough at finding software flaws to be a national-security concern, is the same capability that reshapes a lot of skilled work. The technology has crossed a line the rules, the governance and the workforce have not caught up to.

On the other side: this is a real step toward the thing the industry keeps promising. Fable 5 is the public face of what Anthropic calls its Mythos-class tier, built for agentic work, models that act rather than just answer, and that can grind for hours on real engineering tasks. The capability that got it banned is the capability that makes it matter. This is the agentic, do-the-work tier, and it is the closest the AGI conversation has come to being practical rather than theoretical. AGI, artificial general intelligence, is the milestone everyone is chasing: a machine that can turn its hand to almost any task a person can, instead of being brilliant at one narrow job and useless at everything else.

The readiness gap

Capability is sprinting. The guardrails are walking.

Before anyone declares a winner

Here is the more useful question: best at what. There is no single best model, and I track this closely. On my benchmark read for June, the lead changes with the job. Fable 5 leads overall capability and tops the agentic board. But Google leads the hardest reasoning, the science questions and coding, and OpenAI still leads professional writing and image generation. So when someone tells you a model is the best AI, the only sensible reply is, at what. Pick by the job in front of you, then prove the shortlist on your own real work.

The direction of travel is still hard to miss, and my view is that Anthropic is running up front. The clearest signal is agentic work: a fortnight ago Anthropic sat outside the top three on the main agentic board, and on 14 June Fable 5 entered at number one. On overall capability it now leads too. For enterprise AI specifically, the agentic, coding, do-the-work tier that businesses actually buy, they have changed the game, and it has mostly come together in the last six months.

The honesty appeals to me as much as the capability. Handed a government order it believed was wrong, Anthropic published the technical detail, said plainly that it disagreed, and complied anyway. In an industry that runs on hype, that kind of transparency is rare, and it is a big part of why I trust their work more than most.

Are we closer to AGI than we think?

On coding, my read is that we are essentially there. The best models now score above 90% on the hardest fresh coding tests, near the ceiling of what those benchmarks can measure. The newer leap is agentic, models that act on a problem for hours rather than just write the code, and that is the capability a government just decided was too dangerous to share. That is not nothing. Still, a benchmark tells you a model is strong at a task, not that it is generally intelligent. My honest answer is that we are closer than the hype-weary crowd assumes on capability, and nowhere near as close on control.

But let us be honest. We are not ready for most of this. The capability is sprinting, and the guardrails, the governance and the jobs conversation are walking. Last week made the point twice over: a model good enough to worry a government, and a model that same government could switch off in an afternoon, taking every business built on it along for the ride.

And notice who got switched off

Every frontier model worth the name is American or Chinese. Look at the chart again: not one UK or European model appears. Europe's best shot, Mistral, competes on sovereignty and open weight rather than raw power, and the UK's crown jewel, DeepMind, is American-owned. So an order that cut every non-American off from the best model overnight exposed the real position. Britain and Europe are not players in this race, they are customers, and their sharpest tools can be switched off by a government that is not their own. If you are building here, treat that as a supply risk, not a headline.

None of that is a reason to sit it out. It is a reason to build with your eyes open. Know which model wins which job rather than betting the house on one. Keep a fallback, including an open-weight model you can run yourself. Own the decision rather than handing it to the cleverest tool in the room. And treat capability and readiness as two different clocks, because right now they are telling very different times.

The close

Capability and readiness are two different clocks.

Fable 5 was the tremor. Whether we are ready is the earthquake. The models are already capable of far more than we are set up to handle, and closing that gap, not chasing the next leaderboard, is the real work.

More insights

AI

Why AI, why now

Leadership

The case for fractional

AI

Never let IT run wild with AI

Ready to move the numbers.

Book a 30-minute impact discovery, no obligation. A clear read on where brand, e-commerce and AI can move growth.