Yuval Avidani
Author
Turns out the most powerful model Anthropic has ever released to the public got stopped by the US government, and this exact week it's coming back to the world. I'm talking about Claude Fable 5, a model that was frozen on June 12th by order of the US Department of Commerce for national security reasons, and after two and a half weeks the brakes were released and it's back in action. This isn't a story about a technical glitch. This is the story of the moment an AI capability crossed a line that made a government treat it like a weapon.
Let's unpack this slowly, because there are several layers here, and each one is interesting in its own right.
What even is Fable 5, and why did everyone get excited
Let's start with the basics. On June 9, 2026, Anthropic announced a new model called Claude Fable 5. Anthropic described it as a model from the "Mythos class" that they made safe for general use. What does that mean in practice: there's a very powerful base model, and Fable 5 is its public version, with guardrails on.
At the same time, Anthropic also released a sibling model called Claude Mythos 5. It's the exact same base model, just with some of the guardrails removed, given to a small group of cyber-defenders and infrastructure providers. Think of it like this: same sports car, but the public version gets a speed limiter installed, while a small professional group gets the key without the limiter.
Now for the numbers, because they tell the story. Fable 5 completed, for Stripe, a full codebase migration in a single day, work that would have taken a manual team about two months. That's what Anthropic quoted: the model "compressed months of engineering into days." For those not in the dev world: "code migration" is like moving apartments with all your furniture at once, dismantling, packing and reassembling everything without breaking anything. Here it happened in a day.
Pricing, in terms of tokens (the units of text the model reads and writes): ten dollars per million input tokens, and fifty dollars per million output tokens. Relatively expensive, and that makes sense for a model sitting at the top.
The guardrails: how does a model become "safe for use"
This is the point that's most important to understand, because this is where all the drama begins. How do you turn a very powerful model into something "safe for the public"? Anthropic built a mechanism: when a certain query comes in that's considered sensitive, the model doesn't answer it itself but routes it to another model, Claude Opus 4.8. According to Anthropic, this routing kicks in on average in fewer than 5% of sessions. Meaning 95% of the time you're talking to the full Fable 5, and only in sensitive cases is there a gatekeeper that sends the question somewhere else, more cautious.
On the capabilities side, Anthropic reported the model reached state-of-the-art level on vision tasks. My favorite example: it finished the game Pokemon FireRed using vision alone, with no access to the game's code, simply by looking at the screen the way a kid would. It also ranked first on Cognition's benchmark called FrontierCode and on Hebbia's finance benchmark. A benchmark, for those unfamiliar, is a standard test run on all models to compare them under the same conditions.
And then the government showed up
So far this is a regular announcement story about a good model. The point that turns it into a headline story is what happened three days later.
On June 12, 2026, the US Department of Commerce ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals, including Anthropic employees who aren't US citizens. Notice how severe this is: not just users abroad, but also employees inside the company itself, purely because of their citizenship. This is treatment usually reserved for military technology, not a commercial software product.
What triggered this? A research report from Amazon claiming a "jailbreak" had been found for the model. A jailbreak, in plain words, means bypassing the guardrails we talked about earlier and getting the model to do things it's not supposed to. In this case, the report claimed the bypass caused Fable 5 to identify security vulnerabilities in code. Picture a model that knows how to read software and point out the holes in the wall, the same holes hackers go looking for.
Anthropic, for its part, downplayed the severity. According to Anthropic, what was actually discovered was "a small number of previously known and negligible vulnerabilities." Meaning not a new cyber weapon, but things that were already known. But as far as the government was concerned, the mere capability was enough to set off a red light.
In my view, this is the most interesting point in the whole story. The model doesn't need to actually do something terrible. It's enough for it to show that it can, for the state to start treating it completely differently.
The lift, and what it cost
The story doesn't end with the freeze. On June 30, 2026, the Department of Commerce completely removed export controls from both models, and Anthropic said it would begin restoring global access starting July 1. In between, around June 26-27, there was an interim phase: guardrails were partially lifted on Mythos 5, and about a hundred organizations regained access, while Fable 5 remained restricted.
Anthropic also confirmed this publicly on its X account: "We've been notified that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow." Short, dry, and it says it all.
But nothing happens for free. Commerce Secretary Howard Lutnick said Anthropic would no longer need an export license, after the company agreed to three conditions. Anthropic committed to proactively identifying security risks, working with the government on standards for future models, and reporting malicious activity to the government. In other words: Anthropic didn't just get its model back, it signed onto a new relationship with the state. That's the price it paid.
Bottom line
In my view, this story is a lot bigger than Fable 5 itself. There's a precedent here. A commercial AI model, a product you're supposed to pay for in tokens and run in the cloud, was treated for a moment under export controls like a fighter jet or a weapons system. The line between "software product" and "national security asset" just got blurry, and fast.
I also want to be honest about the limits of what we know. We know the skeleton: what was announced, what was frozen, when it was lifted, and what Anthropic committed to. We don't know how severe Amazon's jailbreak actually was, because Anthropic says "negligible" and the government treated it as a threat. That gap hasn't fully closed for us, and it's worth remembering that.
And the distinction that bothers me the most: Fable 5 is the version with guardrails, for the public. Mythos 5 is the same model without some of those guardrails, for a select group of cyber-defenders. Someone decided what's safe enough for us and what's reserved for professionals only. So here's the question I'll leave you with: who should actually decide what's "too safe" to release to all of us, the company that built the model, or the government that can shut it down?