Three weeks ago I wrote here that they had switched off Fable 5: the most capable AI model available to the public, disabled for everyone by a US government order. I also gave it a podcast episode, The leash pulled. The thesis was uncomfortable: when you build on top of one of these tools, you depend on something a third party can switch off whenever it wants.

Well: today, July 1st, Fable 5 is back. And the outcome, on the face of it, is a happy ending. But if you look calmly, it doesn’t disprove the lesson of the shutdown. It underlines it.

What happened, without dressing it up #

On June 30th, the US Commerce Department lifted the export controls it had placed on Fable 5 (and its restricted sibling, Mythos 5) nineteen days earlier. The next day, today, Anthropic switched the model back on globally: on Claude.ai, on the platform, in Claude Code and in Claude Cowork. For this first week it comes back with usage capped at half the normal limits, and full availability returns from July 7th.

Let’s recall the reason for the shutdown: a jailbreak —a way to bypass the safeguards— that Amazon researchers found in Fable 5, related to analyzing code for vulnerabilities. The government treated it as a national security risk; Anthropic argued that capability already exists in other public models.

It’s back, but with a new leash #

And here’s the part that interests me. It hasn’t come back unchanged. It’s come back more closely watched:

• A new classifier blocks that specific jailbreak in more than 99% of attempts.
• Any request that trips that alarm is automatically rerouted to Opus 4.8, with a notice to the user. That is: there are requests that, even if you think you’re talking to Fable 5, get served by another model without you deciding it.
• And, as part of the deal to reactivate it, Anthropic commits to detecting and reporting malicious activity to the government and to helping shape the standards for the next models.

None of this is unreasonable —they’re sensible safety measures—. But the summary is clear: the tool comes back, yes, and it comes back with more layers of control above you. The leash hasn’t come off. It’s been loosened a bit, and a link has been added.

What I really take away #

Three weeks ago I wrote that the leash “you don’t see it when everything’s fine; you feel it the day it pulls”. This month has shown the whole rope, in both directions.

Because the striking thing isn’t just that they switched it off overnight. It’s that they switched it back on overnight too, again by a decision that doesn’t depend on you or on the company that makes it, but on an office above them both. Nineteen days. A model someone could have built their work on was dead for nineteen days and came back to life through an administrative change. The availability of your most powerful tool, subject to the timelines of a ministry.

The happy ending erases nothing: it confirms that the switch isn’t yours. It doesn’t matter that this time they flipped it in the good direction.

And yes, I’m still married to Anthropic #

I say it again because it would be dishonest not to: I’m the first one who doesn’t practice what I preach. I use these models every day, I have my workflow built around them, and the whole episode —shutdown and switch-back-on— is the textbook case of everything I’ve been saying for weeks… and even so, here I am, with no self-hosted plan B.

Maybe that’s the most uncomfortable lesson of the three weeks: not that the tool can be switched off, which we already knew. It’s that it can be switched off, leave you without it for nineteen days, come back as if nothing happened, and you’re still exactly where you were. Habit weighs more than the scare.