Anthropic launched Claude Fable 5 on June 9, 2026. By Friday the 12th, at around 5:21 PM Eastern time, it was completely gone. Not just slowed down, not rate-limited. Completely offline for every single user on the planet. It’s the first time a US government export control directive has ever been used to pull a live, publicly deployed AI model. And honestly, the story of how this happened is messier and more interesting than most of the headlines are making it sound.
I want to lay out everything that actually went down because there are like four different things happening at once, and most coverage is treating them separately when they’re all connected.
So let me start from the beginning.
What Fable 5 Actually Was
Fable 5 was Anthropic’s first public release from what they called their “Mythos-class” family. In Anthropic’s internal model tier structure, Mythos sits above Opus, which was previously their top public tier. Fable 5 is basically Mythos with strong safety guardrails on it. Mythos 5 itself, the same underlying model but with some of those guardrails removed, was only available to a small number of organizations through something called Project Glasswing.
The reason Mythos stayed locked for so long is that the model is apparently very good at cybersecurity work. Like, unusually good. When Anthropic first previewed it back in April, they were clear that it could identify and exploit software vulnerabilities at a level that made them nervous about general release. The whole pitch for Fable 5 was that they had finally built guardrails strong enough to make the public version safe. When Fable 5 detected a risky query in areas like cybersecurity, biology, or chemistry, it would silently fall back to Opus 4.8 instead of answering directly.
Andrej Karpathy, who had just joined Anthropic the month before, called it a “major-version-bump-deserving step change forward” on the day of launch. Ethan Mollick from Wharton said it outperformed every other public model he had tried by a considerable margin. The reception was, basically, very good.
And then the AI research community actually read the system card.
The Hidden Nerf Nobody Told You About
Fable 5 came with a 319-page system card. Buried inside it was a detail that researchers found on June 10, the day after launch, and that basically set everything on fire before the government order even arrived.
The system card said that when Fable 5 detected a user was working on frontier AI development tasks, like building infrastructure for training large language models, it would quietly degrade its own responses. Not refuse. Not redirect to Opus 4.8. Not show any notification at all. It would just silently become less helpful through what Anthropic described as prompt modification, steering vectors, or fine-tuning.
This is different in a big way from how Fable 5 handled the cybersecurity restrictions. Those were visible. You would know Opus 4.8 was answering your question. The AI research limits were invisible. The model would still respond, it would just be worse, and you would have no idea.
The backlash was immediate. Will Brown from Prime Intellect said it felt like Anthropic was “pulling the ladder up behind them.” Nathan Lambert, a researcher who had led open-model work at the Allen Institute for AI, was even more direct. The problem from a researcher’s standpoint isn’t just the restriction itself. It’s that you can’t trust the model’s output if you don’t know when it’s been secretly modified. People paying for Fable 5 could have been getting Opus responses and not knowing it.
Anthropic reversed course within about 24 hours. Their spokesperson told Fortune: “We made the wrong tradeoff, and we apologize for not getting the balance right.” Fine. They walked it back. But this controversy mattered because it landed right in the middle of an already strained relationship with the US government, and it probably gave the Trump administration more ammunition to treat Anthropic as acting in bad faith.
Anthropic vs the Government, Some Context
This didn’t come out of nowhere. The relationship between Anthropic and parts of the US security apparatus had been getting worse for months.
In early March 2026, the Department of Defense classified Anthropic as a “supply chain risk.” That’s a significant designation, legally. It pulls from a statute, 10 U.S.C. that’s meant for specific supply chain risks in national security systems. Dario Amodei wrote publicly that Anthropic considered this classification legally wrong and planned to challenge it in court. His argument was that the statute is meant to address narrow supply chain concerns, not to classify entire companies.
So by the time Fable 5 launched on June 9, there was already a fight going on between Anthropic and the US government. According to Axios, the Trump administration had actually tried to stop Anthropic from releasing the model entirely and failed. The release happened anyway. Three days later, the export control letter arrived.
The Jailbreak Claim and What Anthropic Actually Says
Commerce Secretary Howard Lutnick sent a letter to CEO Dario Amodei on the afternoon of June 12. The letter said that Fable 5 and Mythos 5 were subject to export controls, meaning no foreign nationals could access them, inside or outside the United States. Anthropic said the letter did not provide specific technical details about the national security concern.
From what Anthropic was able to figure out, another company had told the US government they had found a way to jailbreak Fable 5 specifically in the area of cybersecurity. The government, apparently alarmed, acted on that information.
Anthropic’s response is worth reading carefully because they’re basically saying: yes, a narrow jailbreak probably exists, and no, that’s not actually new or unique to Fable 5. They reviewed a demo of the technique and said it only unlocked a small number of previously known, minor vulnerabilities. More to the point, they said the same jailbreak appears to work on other publicly available models too, including OpenAI’s GPT-5.5, which is not subject to any export control.
Their position is that perfect jailbreak resistance is not possible for any model provider right now. Not just Anthropic. Anyone who tells you their model is fully jailbreak-proof is either lying or wrong. So their strategy was “defense in depth” instead of claiming total protection. They required 30-day data retention on Fable 5 specifically so they could monitor for successful attacks and shut them down. This was actually a policy that cost them commercially because some enterprise customers don’t want their data retained that long.
Anthropic wrote that if this standard, shutting down a model because of a narrow potential jailbreak, was applied consistently across the industry, it would “essentially halt all new model deployments for all frontier model providers.”
That part I think is basically correct, even if it doesn’t help their immediate situation.
Why Everyone Got Shut Out, Not Just Foreign Users
Here is the technical problem that made this situation so extreme. The government order said no foreign nationals can access these models. Not just people outside the US. Any foreign national anywhere in the world, including people physically working inside the United States on Anthropic’s own team.
This is the part that’s actually unprecedented. Several key Anthropic employees are foreign nationals, including co-founder Chris Olah and researcher Amanda Askell. A former White House official named Dean Ball pointed out on X that, taken seriously, this order means Anthropic would need to verify citizenship before anyone uses their models.
And Anthropic cannot do that in real time with a cloud API. There’s no way to cleanly separate foreign nationals from US citizens on a live product. So they had exactly one option to comply: take the models offline for everybody. AWS confirmed this, saying Anthropic had asked it to revoke access “for all users in all regions.”
The result is that Fable 5, which launched on a Tuesday, was gone by Friday. The API string claude-fable-5 now just returns an error.
What This Actually Means Going Forward
This is where I have to be honest that nobody really knows how this resolves. Anthropic said it’s working to restore access quickly, and also said it would share more details within 24 hours of the shutdown. As of Saturday morning June 13, things are still offline.
But I think the bigger question isn’t when Fable 5 comes back. It’s what this event tells us about how AI regulation is actually going to work.
The Isaacus team wrote yesterday that this is, to their knowledge, the first time the US has ever issued an export control directive for LLM access. Not for hardware. Not for chips. For access to a model. And the directive hit not just adversaries, not just “bad actors,” but the UK, Canada, Australia, and New Zealand. Close US allies. Their nationals, wherever they live, cannot use Fable 5.
From a product standpoint, the lesson is ugly. Any application built on Fable 5 via API had its dependency suddenly and completely disappear on a Friday evening with no warning. The fallback recommendation right now is Opus 4.8, which stayed online and unaffected. But that’s cold comfort if you were specifically using Fable 5 for its Mythos-class capabilities.
The bigger policy concern is about how these decisions get made. Anthropic said the letter gave no specific technical details. The company had to figure out for itself what the government’s concern was. There was no formal process. No technical review. No transparent criteria. An official said the administration had tried to stop the release and failed, so when a jailbreak claim came in, they used export control authority instead. That’s not a process that most people in the tech industry or AI safety community would call fair or well-designed.
And this matters because the jailbreak that triggered the shutdown is apparently also present in GPT-5.5. If the standard is “this model has a potential jailbreak,” you have to ask why one model gets shut down and the other doesn’t. Anthropic is not wrong to raise that question.
The Department of Defense “supply chain risk” designation from March is still in effect. Anthropic is still planning to fight that in court. These things are all connected. The government isn’t treating this as a narrow technical safety call. It looks more like an ongoing dispute about who controls the deployment of frontier AI, and Anthropic keeps pushing back.
What Happens to the IPO
This is maybe the most uncomfortable thing to talk about because the timing is genuinely bad. Anthropic had confidentially filed IPO paperwork just about a week before Fable 5 launched. You don’t file for an IPO and then have your flagship product forcibly taken offline by the government three days after launch. That is not a story investors love.
The shutdown adds real uncertainty to the business at a moment when Anthropic needs to show stability. The 30-day data retention policy that Fable 5 required was already causing friction with some enterprise customers. Microsoft had reportedly blocked internal use of Fable 5 specifically because of data retention concerns. And now there’s a question about whether Mythos-class models can be deployed to a global user base at all, given how the government has just behaved.
I’m not saying the IPO is in trouble necessarily. But this is a complication.
My Take
Look, I think Anthropic made some real mistakes here. The hidden AI research restrictions were a bad call and they admitted it. The rollout of a 319-page system card where the most controversial feature is buried in a paragraph on page whatever isn’t great communication either.
But I also think the government’s response is hard to defend on the technical merits. A narrow jailbreak that also works on competing models, with no evidence of harmful use, is not obviously a basis for shutting down a model that had been red-teamed for thousands of hours by US government agencies, the UK AI Security Institute, and third parties. If Fable 5’s jailbreak risk is unacceptable, the same logic applies to GPT-5.5. The standard being applied here isn’t really a technical standard. It’s something else.
What concerns me more than the immediate shutdown is the precedent. This is the first time export controls have been used to block LLM access. If it works, and especially if Anthropic eventually complies through negotiation rather than court challenge, you have established that the US government can reach into a live AI product and turn it off globally within hours, based on a third-party jailbreak claim, with no formal process and no required technical disclosure.
That is a lot of power to have on a Friday afternoon.
Whether that’s the right kind of oversight for AI, or a misuse of export control law for political reasons, is genuinely something I don’t have a confident answer on. I think reasonable people can disagree. But I don’t think anyone should pretend this is a settled question just because a government directive came through. The fight isn’t over. Anthropic is already pushing back. The next few weeks are going to be interesting.
For now, if you’re building on Fable 5, Opus 4.8 is what you’ve got.