Summary

  • Claude Fable 5 is scheduled to go back online globally on Wednesday, almost three weeks after U.S. export controls compelled Anthropic to take it and its sibling model Mythos 5 offline.
  • A new safety classifier developed by Anthropic effectively blocks the reported jailbreak method in over 99% of instances.
  • In collaboration with Amazon, Microsoft, and Google, Anthropic is formulating a jailbreak-severity framework and has pledged early government access to future frontier models prior to their release.

Claude Fable 5 will be reintroduced to users worldwide on Wednesday via Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, as the U.S. Commerce Department has lifted the export restrictions that had kept it offline since June 12.

However, the more advanced sibling model, Mythos 5, will remain restricted to select partners only.

In a blog post from late Tuesday, Anthropic stated, “On Friday, June 12, the U.S. government imposed export controls on our latest models, Claude Fable 5 and Claude Mythos 5. This necessitated us to limit access to foreign nationals, whether inside or outside the United States. As of today, June 30, the export restrictions on Fable 5 and Mythos 5 have been lifted.”

The shutdown was prompted by a jailbreak discovery by Amazon researchers, revealing a method to bypass Fable 5's protective measures and identify software vulnerabilities. In one instance, the model generated code that illustrated how to exploit a vulnerability. Given that the export order was applicable to all foreign nationals globally and Anthropic lacked a mechanism to verify user nationality in real-time, the company opted to disable both models for all users to avoid potential noncompliance.

We’ve been informed that the Department of Commerce has lifted export restrictions on Claude Fable 5 and Mythos 5.

We will start restoring access tomorrow, and will provide an update shortly.

We appreciate our users’ patience and all who collaborated with us on…

— Anthropic (@AnthropicAI) June 30, 2026

Anthropic noted that its solution involved retraining a safety classifier specifically designed to detect the reported jailbreak technique. The company claims it now successfully blocks that method in over 99% of attempts, with any flagged requests for Fable 5 automatically redirected to Claude Opus 4.8, accompanied by a user notification.

The Commerce Department's Center for AI Standards and Innovation evaluated both the new and previous safeguards, deeming them exceptionally robust, although Anthropic acknowledged that the stricter filter may also result in more false positives.

Essentially, the already stringent and overly cautious Fable will be even more vigilant by design.

Access specifics will differ by subscription plan. Pro, Max, Team, and select Enterprise users will have Fable 5 included for up to half of their weekly usage limits until July 7, after which it will transition to usage credits, resulting in a reduced window of access—six days at a 50% limit compared to the original 13 days of full access. In the end, users will experience less complimentary access than they would have without the ban, which also aids Anthropic in saving on computing costs.

Availability on AWS, Google Cloud, and Microsoft Foundry will be implemented "as quickly as possible," according to the company, meaning cloud-hosted enterprise deployments will take longer than Claude's native applications.

The downtime allowed competitors to gain an edge. While Mythos and Fable were offline, OpenAI's GPT-5.5-Cyber surpassed Mythos on CyberGym, a benchmark from UC Berkeley that evaluates AI agents against 1,507 known vulnerabilities across 188 open-source projects and rates them based on their ability to reproduce these vulnerabilities.

GPT-5.5-Cyber achieved 85.6%, outperforming Mythos 5’s score of 83.8%, a gap that Anthropic could not immediately address since its model was not operational. Chinese companies also seized the opportunity to promote their own vulnerability detection systems, with Qihoo 360 launching Tulong Feng and Z.ai introducing open-weight versions.

Additionally, smart routing alternatives using less powerful LLMs also demonstrated comparable results to those reported by Mythos, with some even surpassing it.

Anthropic is now suggesting a collaborative approach to evaluate future jailbreaks with Amazon, Microsoft, and Google, assessing them based on capability gain, scope, ease of weaponization, and discoverability. The company has also committed to providing earlier access to the government for testing frontier models and their safeguards before public release, and has initiated a new HackerOne program for researchers to report techniques for bypassing Fable 5’s protections.