Fable 5 is back today, nearly three weeks after it vanished. On June 12 the government applied export controls to Anthropic's two newest models, and because there was no way to check a user's nationality in real time, the company pulled Fable and Mythos for everyone, foreign or not. The controls came off on June 30. As of this morning Fable 5 runs again on Claude.ai, Claude Code, and the rest. That's the headline, and it's the least interesting part of the announcement.

What tripped the whole thing was a bypass Amazon's researchers found: prompt Fable a certain way and it would list software vulnerabilities. Anthropic's own account is unusually deflating about it. The technique "did not expose any unique Mythos-level cyber capabilities," they write, and less capable models could be coaxed into the same behavior. So a model serving hundreds of millions of people got switched off worldwide over something the vendor itself now grades as borderline, and that a weaker model could already do. An afternoon letter, a planet-sized consequence.

The classifier is where the fix lives. A new one blocks the specific technique in "over 99%" of cases, and when it trips, the request quietly reroutes to Opus 4.8. The pricing is its own tell: on Pro, Max, Team and select enterprise plans, Fable is included for up to half your weekly usage through July 7, then switches to paid usage credits. It comes back for a week at the old deal, and after that it comes back for money.

The bigger move is the machinery around it. Anthropic is proposing a shared framework for scoring how severe a jailbreak is, rated on capability gain, breadth, ease of weaponization, and discoverability, built jointly with Amazon, Microsoft and Google. It was Amazon's researchers who found the flaw that pulled Fable in the first place, and Amazon now helps write the scale that decides how bad the next one is. Next to that framework sit four commitments to the government: pre-release access to models and safeguards, faster sharing of jailbreaks, dedicated joint research, and a "voluntary" industry security standard. I called this shape two weeks ago: not a license, because a license implies criteria a court could test, but something that behaves like one while keeping the discretion intact. Here it is, written into a redeployment blog as good citizenship.

I don't think Anthropic is the villain here. A company handed a Friday-afternoon shutdown did roughly what it could to get its product back on. But "back" is carrying a lot of weight in the coverage. The model that returns today isn't the model that left. It left as a thing you could just use. It comes back as a thing you're cleared to use, on terms the labs will help write while the government keeps the pen. Fable launched with a trapdoor. It reopens with a turnstile.

Sources: