Anthropic announced on Tuesday, July 30, the return of Claude Fable 5, its AI model that had been suspended for nearly three weeks following export restrictions imposed by the U.S. government.
The model will be accessible again starting Wednesday, July 1, across all of the company’s platforms (Claude Platform, Claude.ai, Claude Code, and Claude Cowork).
What does this mean in practice for users? For Pro, Max, and Team subscribers, as well as some Enterprise plans, Fable 5 will be included up to 50% of weekly usage limits through July 7, before switching to a credit-based system.
Access via AWS, Google Cloud, and Microsoft Foundry is expected to follow “as soon as possible.” Mythos 5, reserved for a select group of partners through the Glasswing program, was restored on June 26 for certain U.S. organizations, with discussions ongoing to expand access to international partners.
Anthropic also announced that it is working with Amazon, Microsoft, and Google to develop a common framework for assessing the severity of AI model jailbreaks, based on four criteria: the gain in capability offered, its scope, the ease with which it can be weaponized, and its degree of public dissemination. The company also announced the launch of a HackerOne program dedicated to reporting security vulnerabilities in Fable 5, as well as strengthened cooperation with the U.S. government on pre-launch testing of future models with high security stakes.
The origin of the suspension
The incident began on June 12. The U.S. government had imposed export controls on Claude Fable 5 and Claude Mythos 5, two models launched three days earlier, on June 9. Unable to verify users’ nationalities in real time, Anthropic had to suspend access to both models for its entire user base, including within the United States. The U.S. government had forced Anthropic’s hand, effectively compelling the company to remove its models.
This decision followed a report by Amazon researchers who had identified a method to bypass Fable 5’s safeguards: by querying it in a roundabout way, the model eventually identified several vulnerabilities and, in one case, produced code demonstrating the exploitation of one of them.
Anthropic, for its part, claimed to have tested the same method on less advanced competing models (including Opus 4.8, GPT-5.5, and Kimi K2.7), which proved capable of identifying the same vulnerabilities. According to the company, the reported technique did not grant access to any cyber capabilities specific to Mythos 5, but rather fell into a gray area of Fable 5’s safeguards related to traditional cyberdefense tasks deemed low-risk.
A new “security classifier” has since been trained in coordination with U.S. authorities to specifically block the technique described in the report, with a reported effectiveness rate of over 99%.
