Three Days After Launch, the Government Pulled Anthropic’s Most Powerful AI Models
What the Fable 5 and Mythos 5 shutoff actually means for anyone building a business on AI tools.
On Friday, June 12, the US government ordered Anthropic to shut off access to its two most powerful AI models. By that evening, Claude Fable 5 and Claude Mythos 5 were gone. Not slowed down. Not restricted. Gone, for everyone, everywhere.
Fable 5 had been available to the public for just three days.
The story has already sparked “the government banned AI” headlines and that’s not what happened. So before you panic or rebuild your whole strategy, let me get you sorted on what this actually was.
What actually got shut down
The order came from the Commerce Department at 5:21pm Eastern. It was an export control directive, the legal lever that governs which technologies can reach foreign nationals, which is why anyone calling this a “government shutdown” has no idea what they are talking about. The government told Anthropic to block Fable 5 and Mythos 5 for any foreign national, whether they were sitting in another country or standing inside Anthropic’s own San Francisco office. (Yes, that includes Anthropic’s own foreign-national employees. Wild right?!?!)
What turned a narrow order into a global blackout is simple. Anthropic can’t reliably tell, in real time, which of its users are foreign nationals. So to comply, it had to pull both models for everybody. Everywhere. Every other Claude model stayed online. Opus, Sonnet, Haiku, all fine. Only the two most capable ones went dark.
A word on what these two even are, because they’re not the Claude most people have been using. Mythos is the most capable model Anthropic has ever built. The company kept it locked down for months because it was unsettlingly good at finding security holes in software. In testing, it found flaws in every major operating system and web browser it was pointed at. Rather than hand that to the public, Anthropic ran a closed program called Project Glasswing and gave it to a vetted group of organizations (Amazon and Google among them) to use for defense.
Fable 5 was the public version. Same underlying model, wrapped in guardrails that route high-risk requests like cybersecurity and biology to a weaker model instead. It launched on June 9 and was immediately the most capable AI the public could touch. Three days later it was pulled off the shelf.
Why they did it, and why Anthropic is fighting it
The government’s stated concern is a jailbreak. A jailbreak is a way to talk the model past its safety guardrails.
On the substance, the case looks thin. By Anthropic’s account, the government gave no written specifics, just a verbal concern about a jailbreak, and a narrow one at that. A universal jailbreak would crack a model open across a whole range of harmful uses at once. This wasn’t that. In plain terms, someone got Fable to read a chunk of software code and point out the flaws in it, and the flaws it surfaced were minor and already known. Anthropic’s response, more or less: every capable model already does this, OpenAI’s GPT-5.5 included, and security professionals use that same trick, pointing AI at their own code to surface flaws, to patch holes before attackers find them every single day.
There are two honest ways to read this. Anthropic’s read is that a minor, already-common capability got treated as a national security emergency, and that if you held every company to that standard, no new model would ever ship again. The government’s read, which it hasn’t fully explained, is that the most powerful model on the planet having any exploitable weakness in the cyber domain is worth stopping first and sorting out later. Both of those can be partly true at the same time. Anthropic is complying while calling the whole thing a misunderstanding, and says it’s working to restore access. Anthropic also has a problem with how the government did this, not just whether it’s right. It argues a block on a genuinely unsafe model should come only through a process it calls “transparent, fair, clear, and grounded in technical facts,” and says this directive met none of that. Reporting suggests the order followed talks between the two sides that had already broken down, so this was less a surprise than the end of a conversation that went bad.
There’s an irony here that people are pointing at, and it’s worth mentioning. Anthropic built its whole identity on being the “careful company,” the one that called Mythos too dangerous to release. Whether or not that reputation is why regulators moved, it’s the climate they moved in. Sam Altman called this kind of positioning “fear-based marketing” back in April. You can disagree with him on plenty and still notice that when you spend months telling the world your tool is uniquely dangerous, the world, regulators included, tends to believe you.
Why this is bigger than Anthropic
Anthropic isn’t publicly traded, so there’s no Anthropic stock to crater on Monday. Most people stop there. That’s a mistake. The money exposed here sits in the ring of public companies wrapped around Anthropic.
Look at who’s tied in. Last November, Nvidia committed to put up to $10 billion into Anthropic and Microsoft up to $5 billion, and in the same breath Anthropic committed to buy $30 billion of Azure compute and up to a gigawatt of capacity running on Nvidia chips. Amazon, already the primary cloud and training partner, built an $11 billion data center just to train Anthropic’s models. That round valued Anthropic at $350 billion, nearly double where it sat two months earlier. So when you ask who’s exposed, the honest answer is Nvidia, Microsoft, Amazon, and Google, which is to say the heart of the entire market.
Those deals are circular. Nvidia invests in Anthropic, Anthropic turns around and buys Nvidia chips, the purchase commitments lift the valuation, and the valuation justifies the next round of deals. The whole ring rests on one assumption, that Anthropic’s frontier models keep scaling into real commercial demand. Mythos is the model that rocked Wall Street in April when it first showed what it could do. So a government that can reach into the center of that ring and switch off the crown jewel, in an afternoon, on verbal evidence, lands somewhere more serious than a broken contract. The dollars already changed hands. What took a hit is confidence. The whole circular bet assumed the government would leave these companies alone to build and sell. On Friday, it didn’t.
Which brings us to the timing, and here’s where I’ll tell you what I actually think. The order landed at 5:21 on a Friday, after the market closed, with sixty-some hours before it opened again. I don’t believe that was an accident, and you shouldn’t either.
This administration has a track record. Watch how the market-moving calls tend to get made and a rhythm shows up. Tariff threats dropped on a Friday or over a weekend, then half walked back by Sunday so the hit is mostly absorbed before anyone can trade on it. It’s happened often enough that even Bloomberg ran a column on the “curious market timing” of it. Friday after the close is when you release something you don’t want the market reacting to in real time. That’s a pattern, and patterns are what I pay attention to.
So no, I don’t buy that a directive killing the most powerful AI on the planet, the one sitting at the center of $15 billion in fresh Nvidia and Microsoft money, just happened to go out at dinnertime on a Friday. Somebody understood exactly what a Tuesday-morning version of this headline would do to the AI trade, and chose to drop it when the market was closed for the weekend.
There’s a tell that settles it either way. Friday timing only shields the market IF this gets cleaned up over the weekend. So watch Monday. If access quietly returns, or they announce a fix before the open, that’s your confirmation the whole sequence was built to keep it off the tape. If it’s still unresolved when the market opens Monday, the Friday timing didn’t help, and the shutdown becomes a story the market reacts to anyway.
And nobody can tell you how they even fix this. The obvious move, block the foreign users and let everyone else back on, is the exact thing Anthropic already said it can’t do. That’s WHY the whole thing went dark instead of just part of it. So there’s no clean path back. The government can reverse the order, which it could do in a day, or Anthropic has to build a way to prove who’s a foreign national that it doesn’t have and can’t build over a weekend. The quick way out is the government backing off. But watch how it backs off. If the models come back and the government never admits it got this wrong, nothing really changed. It just proved it can reach in and switch off any model it wants, whenever it wants.
The part that actually matters for you
I used Fable for the first time yesterday, and it was mindblowing. I had it build landing pages, and it turned them out in a fraction of the time Opus takes, at a quality that wasn’t even close. Then I handed it ten system setup and structure documents from a client and asked it to evaluate them. In ten minutes it walked me through all ten. What was solid, what needed work, and what straight up needed to be rebuilt from the ground up. Ten minutes. That kind of review used to eat my whole afternoon.
And then it was gone.
The reason this matters isn’t that I lost a shiny new toy for a day. Almost nobody was running on Fable. It was three days old, and the only people who had their hands on it were the ones who live in this stuff, like me. That’s exactly what makes the shutoff worth your attention. The most capable AI on the planet, with the biggest companies on earth and $15 billion behind it, got switched off by one letter in an afternoon. If that can happen to that, nothing you build on is as solid as the marketing makes it look.
This isn’t a “don’t put all your eggs in one basket” lesson either, because spreading your work across three models wouldn’t have saved you here. The order took out the whole top tier at once. This time the threat came from a government, and you can’t switch your way out of that, because the next model sits under the same letter. The only thing that sits outside all of it is the part of your business that was never a tool to begin with. The people who trust you. Your own judgment about what’s worth saying and selling. Build those like they’re the asset, because they’re the one thing no order can switch off. Use every AI tool you can get your hands on, hard, the way I do. Just don’t confuse the tool with the business.
Where this leaves you
The doom crowd will tell you this is the government coming for AI. Maybe. The truth is we don’t know yet what’s going to come of this. The world of AI is changing at light speed almost daily. For now, this is just one government order, built on a jailbreak claim Anthropic says doesn’t hold up. Both sides sound like they want it gone fast.
So don’t rebuild your whole strategy on a story that’s still being written. Stay clear-eyed, keep building on ground you actually hold, and watch how this resolves, because the how matters more than the headline.
The most powerful AI in the world went dark in an afternoon. That’s worth understanding clearly. It’s not worth losing your head over.
If you want the clear version of stories like this one, without the panic and without the hype, this is what I do here every week. Subscribe and I’ll keep you informed.