Anthropic releases Claude Fable 5, a restricted version of its powerful Mythos AI model, amid safety and data concerns

Anthropic launched Claude Fable 5 on 9 June, making its most powerful class of AI models, Mythos, available to the public for the first time with significant restrictions. The model is designed to automatically downgrade responses on sensitive topics like cybersecurity, biology, and chemistry to a less capable model, Claude Opus 4.8, to prevent misuse.

A dual-track safety mechanism

Fable 5 operates on a dual-track system. For standard queries, it uses the full capabilities of the new model. When a user asks about a sensitive subject, the system reroutes the query to Opus 4.8, which has lower performance but is considered safer. Anthropic states this happens in about 5% of sessions, though researcher Ethan Mollick, who had early access, noted that the safety mechanisms trigger "at the slightest hint of a security problem, and it happens far too often."

Fable's safety mechanisms trigger at the slightest hint of a security problem, and it happens far too often.

The Mythos family and its capabilities

The Mythos family was first unveiled in April and kept restricted due to its ability to find thousands of critical vulnerabilities, including flaws in all major operating systems and browsers. On ExploitBench, a benchmark measuring offensive cybersecurity capability, Mythos 5 scores 78%, compared to 40% for Opus 4.8 and 69% for the earlier Mythos Preview. Alongside Fable 5, Anthropic also officially launched Claude Mythos 5, an unrestricted variant for government agencies and critical infrastructure security operators, available only to a limited set of authorized partners.

Data retention triggers Microsoft restrictions

Microsoft is limiting employee use of Claude Fable 5 due to new data retention requirements. Unlike other Claude models that operate under Zero Data Retention rules, Fable 5 requires Anthropic to retain prompts and outputs for 30 days to operate its safety classifiers. Data flagged for policy violations can be stored for up to two years. Microsoft's legal teams are evaluating these changes, and the model is not available in the internal model picker for GitHub Copilot, even though it has been rolled out to external customers.

European regulators still evaluating access

The European Union Agency for Cybersecurity (ENISA) has not yet accessed the Mythos model, despite an offer from Anthropic to the European Commission on 31 May. A spokesperson said terms and conditions for access still need to be agreed upon, with another meeting between ENISA experts and Anthropic expected next week. ENISA noted that Mythos is not an isolated case and that a new wave of models is arriving on the market.

ENISA is in talks with other Frontier Models. ENISA's priority now is to continue collaborating with Anthropic to agree on the terms and conditions for future access.

Criticism over timing and messaging

Anthropic published a document on 4 June, signed by research head Marina Favaro and policy head Jack Clark, arguing for the option to slow down or temporarily suspend frontier AI development. Five days later, it released its most powerful public model. Critics, including University of Washington linguist Emily M. Bender, describe this pattern as a form of promotion where the alarm functions as a certification of power.

AI doomerism is also AI hype.

Pricing and performance

Via API for enterprise customers, Fable 5 costs $10 per million input tokens and $50 per million output tokens, between 67% and 100% more expensive than OpenAI's GPT-5.5. Stripe reported that Fable 5 compressed months of work into days, executing a migration of 50 million lines of code in a single day, an operation that normally requires over two months of manual work.