Claude Fable 5 is back online: the government lifts the shutdown 19 days later

On June 12 the US government ordered Fable 5 and Mythos 5 switched off for the entire planet. Today, July 1, Fable is answering again. Here's how the ban got lifted, what Anthropic had to give up in return, who was behind the jailbreak that started it all, and what lessons the first federal shutdown of a commercial AI leaves behind.

Published on 1 July 2026

The end (for now) of the saga

If you've been following the blog these past weeks, you know the story: on June 9 Anthropic launched Claude Fable 5, the AI that finds 27-year-old flaws in your kernel; three days later the US government ordered it shut down along with its bigger sibling Mythos; and while it was dark we shared the lesser-known facts about the model nobody was allowed to touch. We closed that second article saying Anthropic was working to restore access and that we'd keep you posted.

Well, here it is: on the evening of June 30 the Commerce Department lifted the export controls, and today, July 1, Fable 5 is available worldwide again. We watched it happen live this very evening: around 9:30 PM Spanish time the model was answering again, and at 9:47 PM Anthropic's official email confirming it landed in our inbox. Nineteen days of total blackout, most of them spent in closed-door negotiations between Anthropic and the government.

As usual, the headlines stop at "it's back." The interesting part is how, in exchange for what, and a couple of details that confirm —point by point— what we'd been warning about. Let's dig in.

The timeline of the unlock

June 26 — the government approves the first step: Mythos 5 comes back online for a set of US organizations in the Glasswing program. The verified defenders get their tool back before anyone else.
June 30, in the evening — the administration announces it is lifting the export controls on both models. Anthropic confirms that same night.
July 1 — Fable 5 becomes globally accessible again: claude.ai, the API platform, Claude Code and Claude Cowork. Reactivation on AWS, Google Cloud and Microsoft Foundry is underway but running a few days behind.

Note the order, because it's no accident: verified defenders first, then the general public, and the hyperscalers last. Exactly the same ladder of trust Anthropic already used for the original launch.

What changed to make the government say yes

Here's the meat. The model coming back today is the same one that went dark on June 12: same brain, same capabilities. What changed is the lock. Anthropic retrained its safety classifier to specifically detect and block the jailbreak technique that started it all, and claims it now stops it in more than 99% of attempts, routing those requests to Opus 4.8 —the same fallback mechanism we already knew—. Researchers from NIST (the US standards institute) validated the safeguards before the unlock.

And there's some fine print anyone who codes for a living will care about: Anthropic itself admits the new classifier, being stricter, fires more false alarms during routine programming and debugging tasks. It's the classic toll of tightening a filter: fewer holes, more false positives. If Fable reroutes you to Opus on a perfectly innocent request these days, now you know why.

The "other company" had a name: Amazon

In the shutdown article we reported that the trigger was "another company" claiming to have found a Mythos jailbreak. It's now public who that was: Amazon researchers —a Glasswing partner, and one of Anthropic's biggest investors, by the way—. They found a way past Fable's safeguards to make it identify vulnerabilities and, in at least one case, generate the code to exploit one. They reported it, the warning reached the government, and you know the rest.

It's a detail that captures the moment we're living in: the defender-collaboration program Anthropic built to showcase its transparency turned out to be, precisely, the channel through which its shutdown arrived. That's the thing about transparency: it works in both directions.

The price of the unlock

This wasn't a "fine, we believe you, switch it on." The deal comes with strings attached that set a precedent for the whole industry:

Pre-deployment government access: models that "materially advance frontier capability" in areas sensitive to national security must be shown to the government before they ship.
Threat-intelligence sharing: fast disclosure of jailbreaks and of how people attempt to abuse these tools.
Participation in a vulnerability clearinghouse created by executive order, alongside Amazon, Google and Microsoft, to jointly classify the flaws AI uncovers.

Anthropic, for its part, has proposed something that —to be fair— was overdue: a common framework to grade the severity of a jailbreak, with four criteria (how much extra capability it grants over existing tools, how many distinct offensive tasks it affects, how much human effort it takes to weaponize, and how easy it is to discover). Its underlying complaint is reasonable: today "there is no agreed-upon standard for classifying the severity of a jailbreak," so the decision to switch off a model deployed to hundreds of millions of people is left to whoever happens to be holding the pen.

Meanwhile, the competition didn't wait

And here's the data point that confirms what we said in the curiosities article: switching off a model doesn't put the capability back in the box. During the blackout weeks, OpenAI shipped GPT-5.5-Cyber and took the cybersecurity benchmark crown from Mythos: 85.6% versus 83.8%. In other words: while the "too dangerous" model sat sealed under federal order, a competitor released an equivalent —or better— version of the exact capability that got it sealed.

We're not pointing fingers. We bring it up because it's the empirical proof of the thesis we keep repeating: this capability already exists in the world, with or without Fable. The only variable you control is whether your infrastructure is patched when someone points it your way.

What actually changes for you

Let's bring this down to earth, because that's what we're here for:

1. If you were using it, you can go back — with a bonus thrown in

Fable 5 is operational again as of today across every Claude surface. Anthropic, by way of apology, is giving paid plans up to 50% extra weekly usage until July 7. If you had workflows half-migrated when the shutdown hit, this week is a good time to pick them back up and decide calmly.

2. The multi-provider lesson is no longer theory: it's 19 measured days

In the shutdown article we told you not to tie production to a single frontier model. That warning now has a number attached: the "most powerful model in the world" was down for nineteen days, for reasons entirely outside its users' control. Whoever had a plan B kept working; whoever didn't spent three weeks staring at an error page. Designing so you can switch providers isn't paranoia: it's exactly this.

3. Assume this will happen again (to this model or another)

The important part of the deal isn't that Fable is back: it's that there is now an institutional procedure for switching commercial AI models off and on, with conditions, timelines and trade-offs. The kill switch has been used once and it worked. Next time —for Anthropic or anyone else— it will be faster and less newsworthy. Factor it into your architecture the same way you factor in that a cloud provider can have an outage.

4. The patching thesis, more alive than ever

With Fable on or off, with Mythos sealed or released, the conclusion hasn't moved an inch through the whole saga: what protects you is not the obscurity of your code, it's your discipline in updating it. There are now at least two frontier models —from two different companies— capable of auditing code at industrial scale. The backlog of latent flaws will keep surfacing, and there's only one way to be on the right side of that list.

The bigger lesson

Three weeks ago we wrote that "what today is the most powerful model ever released can be an error page tomorrow." Today we get to finish the sentence: and the day after tomorrow it can be a product again, with new conditions and a regulator looking over its shoulder. The ground under these tools moves fast in both directions, and the only sensible strategy is still the boring one: try things without marrying them, design so you can switch, and never let a piece of your production depend on a switch you don't control.

We'll keep doing what we've done all along: testing, watching, and telling you about it without the marketing gloss. And if these weeks' blackout got you thinking about how many switches you don't control are wired into your infrastructure —models, APIs, cloud providers—, that's an exercise worth doing properly. If you'd like, we'll do it with you.

References

Want to talk about your case?

Tell us what you need and we will get back to you within 24 hours with a clear proposal.

Get a quote