Anthropic said it has restored access to its frontier models Claude Fable 5 and Claude Mythos 5 after the US government lifted its export controls on the advanced large language model (LLM).

Fable 5 will be accessible globally via all of Anthropic’s platforms including Claude.ai, Claude Code, and Claude Cowork. Until 7 July, users on the Claude Pro, Max, Team, and premium Enterprise plans can use up to half of their weekly usage limits on Fable 5.

After this date, Fable 5 will only be available via usage credits priced at $10 per million input tokens and $50 per million output tokens.

Not all of Fable 5’s capabilities are returning, however. Anthropic said that for the time being, when users ask the model to complete coding or debugging tasks it will instead route to Anthropic’s next-best model Claude Opus 4.8

The US Department of Commerce ordered Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals at the start of June, a move Anthropic said was prompted by Amazon researchers discovering Fable 5 could be used to identify software vulnerabilities and in one example generate code to exploit a flaw.

Rather than restrict access by nationality, Anthropic pulled the models entirely while it assessed the risks.

Both Fable 5 and Mythos 5 are Anthropic’s most advanced LLMs. Mythos, designed specifically to find vulnerabilities, was released to pre-approved organisations in April as part of the Project Glasswing cybersecurity initiative.

The model’s launch prompted widespread discussion about the potential for AI to autonomously detect and exploit cyber vulnerabilities.

Fable 5 draws on the same underlying architecture as Mythos 5 but comes with heavy domain restrictions and guardrails that prevent it from being used for cyber capabilities.

The firm has now said its tests have determined this behaviour is not unique to Fable 5, and that “many less capable models” can identify vulnerabilities in the same manner including Claude Opus 4.8, OpenAI’s GPT-5.5, and the open-source model Kimi K2.7, developed by Chinese firm Moonshot AI.

It added that none of the models, including Fable 5, were able to expose vulnerabilities to the same degree as Mythos 5 and that every AI model it tested was able to generate instructions on how to exploit a single vulnerability.

To address government concerns and prevent users from jailbreaking Fable 5, Anthropic has reworked its ‘classifiers’, automated systems that detect whether the user is asking the model to act in a harmful way.

Anthropic said Fable 5’s classifiers now have an even larger safety margin, blocking harmful requests and even triggering for some prompts that it knows are benign. When a prompt is blocked, users will be shown a notice and the request will instead be routed to Claude Opus 4.8.

“Our classifiers make successful jailbreaks very costly and high effort to produce, and even if a jailbreak is successful, our extra layers of defence provide additional mitigation,” Anthropic wrote in a blog post.

“We’ll continue to update our classifiers as we learn more about novel jailbreak techniques.”

Anthropic also announced programme with the cybersecurity vulnerability platform HackerOne, through which security researchers can report Fable 5 jailbreaks.

---