Claude Fable 5 and Claude Mythos 5 sit among the most capable models Anthropic runs. For about nineteen days this June, neither was available to anyone, and the story of why is worth a firm's attention.
This was not a routine service problem, the sort that resolves itself in an hour with an apology and a status page. Access to two frontier models was withdrawn by order, and it stayed withdrawn until a government department chose to reverse the decision. A firm that reads this as somebody else's news is missing the part that bears on its own practice.
What happened
On 12 June 2026 the US government placed export controls on both models. The trigger was a security finding rather than a commercial one. Researchers found a way to bypass the safeguards on Fable 5, a jailbreak that prompted the model to identify a software vulnerability and, in one case, to produce code showing how that weakness might be exploited. The order took effect at once.
Because Anthropic had no way to verify every user's nationality in real time, and because the controls turned on who was using the tool, the company suspended access for everyone rather than risk breaching the rule. That is the detail that turned a targeted order into a total outage. A tool that had been working one day was gone the next, for reasons that had nothing to do with the firm using it.
How it was resolved
The suspension ran for about nineteen days. On 30 June the US Department of Commerce lifted the controls. Fable 5 returned globally on 1 July, and Mythos 5 was restored to approved organisations. Access came back, but only after a delay that no user had any control over.
It is worth sitting with the shape of that timeline. The order landed on 12 June and took effect the same day, so there was no window in which a firm might have wound down its reliance in an orderly way. The reversal came eighteen days later, and even then it was staged, with Fable 5 open to all and Mythos 5 held back to approved organisations. From the point of view of a firm mid-matter, the tool was there, then it was not, then it returned on a schedule set entirely by others.
Why this matters for a firm
Frontier models now sit close enough to national security policy that a government decision can remove a working tool overnight. This was not a billing dispute or a service outage of the ordinary kind. It was a genuine safety finding that led a regulator to act, and the response reached every user of the tool regardless of where they were or what they were doing with it.
The safety finding is the part that should stop a firm from dismissing this as an American quarrel with no bearing on Manchester or Leeds. A researcher showed that the model's guardrails were possible to work around, and the consequence of that finding fell on lawful users who had done nothing wrong. Where the reason for an outage is a real weakness in the tool rather than a contract dispute, the vendor cannot promise it will not happen again, because the vendor does not control the regulator's response.
For a firm that had built one of these models into its daily work, the lesson is uncomfortable. The reliability of an AI tool depends on more than the vendor keeping the servers running. It depends on a policy environment that can shift without warning. A firm should treat that as a real risk and plan for the day a tool it relies on becomes unavailable, before that day arrives.
Anthropic’s newsroom carries the official account of the suspension and restoration.
If one model going dark would stall your firm's work, the continuity plan needs writing before the next suspension, and we write them: tell us where you stand.