Yuval Avidani
Author
"The smartest public model in the world was available for three days, and then a government pulled it offline." This isn't a scene from a thriller, it's literally what happened this month with Claude Fable 5.
Let's unpack this slowly, because there are layers here. On June 9, 2026, Anthropic released Claude Fable 5, and instantly it was the most advanced public model that existed. Not "one of the best", the best. And then, within just three days, the US Secretary of Commerce ordered it shut down for every foreign national in the world. Including people physically working inside the US. The most powerful model in the world lived exactly three days before regulation ran it over.
What Claude Fable 5 actually is
Let's start with the basics, because without this you can't understand why this story is such a big deal.
Claude Fable 5 is a large language model from Anthropic built specifically for long-horizon tasks. Think of it this way: most models are good at answering one question and wrapping things up. A long-horizon task is a completely different animal, it's like handing an employee an entire project that takes hours or days, with dozens of steps depending on each other, without losing the thread halfway through. That's the difference between a model that answers our question in a chat window versus a model we can leave alone with a complex task, and come back to a finished result.
And here's the number that made the noise. On a benchmark called the Artificial Analysis Intelligence Index, an independent, external benchmark that ranks models by how "smart" they are across a range of tests, Claude Fable 5 scored 60. For perspective: Opus 4.8 scored 56, and GPT-5.5 scored 55.
Four points sounds like nothing, but in the world of language models, at these levels, every single point is a battle. A four-point gap at the top of the chart isn't a nuance, it's a distance you can feel. It put Fable 5 alone at the summit, at least for a moment.
And then came the letter
The wild part, and in this case also the unsettling part, is what happened next.
Within three days of launch, US Secretary of Commerce Howard Lutnick sent a letter to Anthropic CEO Dario Amodei. The instruction was unambiguous: disable access to Fable 5, and its sibling model Claude Mythos 5, for every foreign national anywhere in the world. Including foreigners physically working inside the US. The official justification: national security.
Let's talk for a second about the legal mechanism here, because it's the heart of the matter. What got triggered is called export control. Think of it like a list of products you're not allowed to sell or move outside the country without approval, exactly like the restrictions on exporting weapons or advanced chips. The administration's logic: an AI model powerful enough stops being just another app, it becomes a strategic asset. So who gets to access it isn't just a business question anymore, it's a security question.
The specific reason Anthropic gave was a jailbreak claim. Let me explain that term, because it's central here. A jailbreak is a break in the guardrails developers put on a model, sort of a backdoor that makes it do things it's not supposed to. Think of it like convincing a disciplined employee, through clever wording, to bypass their own protocols and do something they shouldn't. In this case, the claim was about a "narrow, non-universal" jailbreak, meaning not a wide breach that takes down all the safeguards, but a specific, narrow hole. And here's the detail that bothers me: the administration only provided verbal evidence for this, not a full technical proof.
Why this is complicated, and why it's fair to look at both sides
Now, I want to be fair here, because it's easy to fall into a "evil government strangling innovation" story. It's not that simple.
On one hand, there's a genuine security argument here. If a model is truly capable of assisting in weapons creation or a dangerous cyberattack, then a regulator can't just say "trust the company." From this angle, export control is exactly the tool countries use for any dual-use technology.
On the other hand, and here's my critique, when the only evidence is a "verbal claim" about a narrow jailbreak, real tension emerges. Because if the mechanism gets triggered based on a claim and not on a full, public demonstration, then the power to shut down the world's most powerful model is standing on very thin ground. Anthropic, by the way, complied. The models went dark.
It's important to clarify the scale of this "takedown." The models were offline for about fifteen days. According to an Axios report from June 27, they were on track to come back within days after negotiations between Anthropic and the administration progressed. In other words, this is a temporary suspension, not a permanent cancellation. It's important not to inflate this into "the model got deleted." It was suspended, and it looks like it's coming back.
The bigger picture: who actually controls AI
Now let's zoom out from this specific event for a second, because it's part of a bigger wave.
Fable 5 wasn't alone. That same wave of export restrictions and national security measures on powerful AI models also hit GPT-5.6. We're seeing a pattern here: as models get more powerful, they're gradually shifting into the regulatory category of "national strategic asset," instead of "regular software product."
And here's the point that fascinates me the most. While these closed models were behind a wall, who filled the vacuum? Open-source models. Let me explain the term: an open-source model, or open-weights model, is a model whose weights, its inner core, are available for anyone to download. Think of it like a recipe published for free: once it's out there, you can't really "turn it off," because it's already sitting on everyone's hard drive.
And that's exactly the twist. When the strongest closed model got shut down by government order, it was precisely open-source, the thing you can't switch off, that filled the void. That says something profound about control over AI: you can rein in one company, but it's a lot harder to rein in knowledge that's already scattered everywhere.
Bottom line, as I see it
So let's sum up in the format I like. Claude Fable 5 is a large language model from Anthropic that specializes in long-horizon tasks, for users and developers who need a model that can carry an entire project from start to finish.
In my eyes, the story here isn't the model. It's the tension. On one side we have a push for power, a race to the top of the leaderboard, who's the smartest. On the other side we have regulation and national security that can shut that leader down within three days, based on verbal evidence alone. And in the middle, open-source, swimming under the wall and not asking permission.
My limitations here: I'm relying on what's been published, and a large part of this story is still unfolding. The jailbreak claim hasn't been fully proven to the public, and the details of the negotiation are partial. I'm not claiming to know whether the move was justified from a security standpoint or not, I'm claiming that the mechanism itself, the ability to shut down the most powerful tool in existence by order, is the thing that should keep us awake.
And finally, the question that genuinely interests me: if real control over AI is shifting from companies to governments, but open-source doesn't get the memo and just keeps spreading, who's actually going to hold the power two years from now?