Anthropic’s Claude Fable 5 Debuts, Microsoft Cautious on Data Terms
Anthropic has launched Claude Fable 5, its first model from the new Mythos class. While the model is now available to GitHub Copilot and Foundry customers, Microsoft has restricted internal employee use pending a legal review of new data retention requirements.
Key Concern: Anthropic now requires prompts and outputs to be retained for 30 days to operate safety classifiers—a shift from the "Zero Data Retention" rules applied to other Claude models.
Data Retention Shift
Under the new policy:
- All prompts and outputs are retained for 30 days, then deleted.
- Content flagged for policy violations may be stored for up to two years.
This change has led Microsoft’s legal teams to evaluate whether Claude Fable 5 can be used for internal projects, particularly those involving customer or confidential data. For now, the model is absent from the internal Copilot model picker used by Microsoft employees.
Background on the Model
Claude Fable 5 is the first broad release from Anthropic’s Mythos class. Anthropic had previously withheld the family from public release, citing its advanced capabilities in cybersecurity tasks. Prompt safeguards were reworked for Fable 5, prompting the updated data retention terms.
Outlook Unclear
It remains uncertain whether Claude Fable 5 will be cleared for internal Microsoft use. Microsoft declined to comment on the ongoing evaluation.