For now, Microsoft is holding back Anthropic’s latest AI technology for internal use. The company is investigating the implications of updated data retention rules in Claude Fable 5, the recently introduced model from Anthropic’s new Mythos series.
Although Microsoft has now made the model available to customers via GitHub Copilot and Azure AI Foundry, its own employees cannot yet use it within internal AI environments. According to a report by The Verge, Claude Fable 5 is currently missing from the internal model catalog used by Microsoft employees. Other Claude models remain accessible, however.
This reluctance is linked to a change in how Anthropic processes data. While earlier Claude models could operate without storing user data, Claude Fable 5 utilizes new security mechanisms that require prompts and generated responses to be temporarily stored. According to Anthropic, that data is deleted after thirty days.
Longer retention periods apply when the system may detect abuse or violations of the usage policy. In such cases, entered data and generated responses may be stored for up to two years. It is precisely this aspect that is reportedly prompting a legal review within Microsoft.
In-house lawyers are currently investigating the implications of the new terms for the model’s use within the organization. According to reports, the focus is primarily on protecting customer information and other confidential data. A final decision on the internal use of Claude Fable 5 has not yet been made.
Most powerful Claude model
Claude Fable 5 marks the first broad introduction of Anthropic’s Mythos family. According to the company, it is also the most powerful model it has made available for general use to date. Anthropic highlights strong performance in software development and analytical tasks, two domains in which Microsoft is also actively positioning its AI services.
The Mythos models previously drew attention after Anthropic indicated that their cybersecurity capabilities were so extensive that additional security measures were necessary before a public release was possible. The company has since built in various restrictions designed to prevent use in high-risk applications.
That extra layer of security appears to be directly linked to the new retention periods. By temporarily storing prompts and responses, Anthropic can perform automated checks for potentially harmful or undesirable use of the model. These rules apply to all platforms on which Mythos models are offered.
Microsoft has not provided a substantive response to the reports.