Anthropic had one hell of a week! On Tuesday, June 9, it released Claude Fable 5, the first publicly available version of its restricted “Mythos-class” technology, and by Friday, the U.S. government had effectively forced it off the market. In between launch and shutdown, the model drew a wave of user backlash over cost (double previous models), transparency (it would switch back to previous models without telling you), and data retention (30 days of mandatory data retention).
According to Decrypt, one test burned through a $100 Max subscription's daily allowance in under nine minutes, and Scrimba CEO Per Borgen reported burning 1.3M tokens in seven minutes, which he calculated at $160 an hour. Anthropic's explanation was that its new “Workflow mode” breaks prompts into parallel subagent tasks that cost more compute by design, and that per-task efficiency looks better than per-token pricing suggests, but users on capped plans experienced it as setting their tokens on fire.
The second major complaint was about how Fable 5's system would quietly degrade its own responses through prompt modification when the model detected the user was working on frontier AI / LLM development, without telling the user. Anthropic estimated that the practice only affected ~0.03% of traffic, but for many users it was more about the principle, not the percentage. After the backlash, the company reversed course, saying it would now tell users when a request was being downgraded and routed to Opus 4.8.
Lastly, there was a mandatory 30-day data-retention policy on all Mythos-class traffic, which Anthropic framed as a security measure, but users noted it effectively locked out GDPR-bound European companies and anyone with zero-retention compliance requirements.
None of those complaints mattered within a few days, when the U.S. government stepped in, and effectively shut down the model.
On Friday, June 12, the Commerce Department issued an export-control directive citing national security, barring any foreign national, including Anthropic's own non-citizen employees, from accessing Fable 5 and Mythos 5. Because so many of Anthropic's researchers are foreign-born and the rule was so broad, Anthropic said it had no choice but to disable both models entirely for everyone.
The Wall Street Journal and others reported that the trigger for the government's decision was Amazon.
CEO Andy Jassy, whose company is both a major Anthropic investor and one of its compute providers, told Treasury Secretary Scott Bessent and other officials that Amazon researchers had used Fable 5 to extract cyberattack-relevant information that was supposed to be off-limits. The administration ran Amazon's findings past the NSA, held a Friday meeting, and reached out to Amodei that same day. Officials asked him to fix the vulnerability or pull the model, which he pushed back on, arguing the bypass was narrow and specific and not a universal “jailbreak.” Shortly after the call, the export controls landed.
The irony of the whole situation was that the same week the government told Anthropic its model was too dangerous to keep selling, CEO Dario Amodei published an essay, “Policy on the AI Exponential,” arguing that transparency rules were no longer enough and the U.S. needed binding, FAA-style safety testing for frontier models, with the power to block or reverse unsafe deployments. Days later, the government did exactly that to his own company.
This is an ongoing, developing story, but hopefully that gives you the gist of what went on last week, and why Claude has a permanently affixed notification above its chat box that reads, “Claude Fable 5 is currently unavailable. Learn more.”
