The Fable case

In June 2026, the US government ordered Anthropic to stop foreign nationals from accessing the Fable 5 and Mythos 5 models. Anthropic had no way to verify nationality, so it shut both models off for everyone. In just under three weeks, the ban was lifted. I had the worry that such a ban would happen before it came. And with it, the question of whether it is even possible to regulate a technology that anyone with rented GPUs and enough money can run and retrain today. Now I have a real case with exact dates for that question.

What happened

On June 9, 2026, Anthropic released the Claude Fable 5 and Claude Mythos 5 models. Fable 5 was thus publicly available for roughly three days, from June 9 until the night of June 12 to 13, 2026.

On June 12, 2026, at 5:21 p.m. ET (11:21 p.m. our time), Anthropic received an export directive from the US government. Verbatim according to Anthropic's statement of June 12, 2026, it required Anthropic to:

suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States

Anthropic responded by shutting off both models for absolutely all customers, not just foreign ones, because by its own account it could not otherwise ensure compliance with the directive:

we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance

The government did not give a reason in the letter. According to Anthropic, what lay behind the directive was the discovery of a way to get around the model's safeguards so that it would search for vulnerabilities in software. Anthropic disagreed with the step, verbatim:

we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people

On June 30, 2026, the government lifted the restriction, as Secretary of Commerce Howard Lutnick announced. From July 1, Fable 5 was available again to users around the world in the Claude apps and through the Claude Platform. It got an improved safety classifier that according to Anthropic blocks the described technique in more than 99% of cases. As of June 30, Mythos 5 returned only to selected US organizations, with government approval from June 26. For Fable 5, the ban lasted just under three weeks. (Source: Redeploying Fable 5, Anthropic, June 30, 2026.)

Where this line of thought came from

Even before the ban, out of fear that it would happen, I wrote down seven points. They are about what usually happens when a government bans access to a powerful and at the same time replicable technology:

A ban does not create safety, it creates demand. The need does not disappear, it just moves elsewhere.
It is not the best that wins, but the most available. A worse model that people actually use gathers both experience and a community.
Speed kills control. Regulation assumes that technology changes more slowly than laws. When development moves in weeks and legislation in years, the regulator is always just reacting to the past.
Protectionism can fund the competition, which takes over the markets left without supply.
The market adapts, but the rules change. New processes and certifications emerge and the regulator becomes part of development.
Competition shifts from the models themselves to who fixes vulnerabilities faster.
Control can turn into surveillance, because registrations and audits create an overview of who uses the technology.

The core is the third point. It is my opinion, not a description of the Fable case. The case just set it against real data for the first time.

Offline yes, cloud no

I had been trying Fable 5 myself since its release. The day before the directive, on June 11, 2026, at 5:53 p.m., I wrote to the model:

I was afraid AI would get banned but I think that's no longer possible. Anyone can have it at home offline now.

I was half right. Nobody banned offline models, and they remained available to people. But the next day, at 11:21 p.m. our time, the export directive reached Anthropic. It turned out that a large model in the cloud can be shut off, for the whole world and within a few hours.

In my view, it was a precedent. There had been national bans before, for example when the Italian data protection authority blocked ChatGPT in Italy at the end of March 2023. Access there came back after a month. But I do not know of a case where a government order led to a frontier model being shut off for everyone in the world. Here it did, even though the government only wanted to cut off foreign nationals. Since June 12, 2026, we know that it can happen.

What the case showed

In my view, the case speaks to the sixth point right away. Fable 5 came back with a new classifier that according to Anthropic blocks the described technique in more than 99% of cases. At the same time, though, Anthropic was showing that weaker models, such as Claude Opus 4.8 or GPT-5.5, find the same vulnerabilities too. What of this decided the return, I cannot tell from public sources.

The third point held in reverse. This time the regulator was not slow. The directive came three days after the model's release and the ban was lifted after just under three weeks. Everything was fast, the ban and its end alike.

For the other points, three weeks was too short. Whether the ban created demand elsewhere, helped the competition, or left behind more lasting surveillance cannot be told from such a short period.

Why the government sent the directive at that particular moment, I know only from Anthropic's version. According to it, the directive came after the government learned of a report by Amazon researchers. They got the model to find vulnerabilities in software and, in one case, even to write a demonstration of how to exploit one of them.

I am left with the question I started with: whether it is possible to regulate a technology that anyone with rented GPUs and enough money can run and retrain today. The Fable case showed that shutting things off is possible, but only for what runs in the cloud at a single company. And this time only for a while, because the government itself lifted the ban within just under three weeks. Whether models from elsewhere would have gotten around a longer ban, I cannot tell from this case.