Microsoft has blocked its own employees from using Anthropic’s Claude Fable 5, the AI company’s newest and most capable model, while its legal team reviews data retention policies that could keep sensitive corporate prompts stored for up to two years.

The restriction, reported on June 10, 2026, creates an awkward split: external Microsoft customers can access Fable 5 through GitHub Copilot and Microsoft Foundry, but the people who actually built those platforms cannot touch the model internally.

The retention problem

Previous versions of Anthropic’s Claude models were available to Microsoft with zero-data-retention configurations, commonly known as ZDR. That meant prompts typed by Microsoft employees and the outputs they received were not stored by Anthropic.

Claude Fable 5, which launched on June 9, 2026, as part of Anthropic’s Mythos-class lineup, changes that calculus entirely. The model mandates a minimum data retention period of 30 days for all prompts and outputs.

If any content gets flagged for potential policy violations, Anthropic’s updated terms allow retention for up to two years. For a company like Microsoft, where employees routinely work with proprietary code, unreleased product details, and confidential business strategy, the idea of a third party holding onto that data for months or years is not exactly comforting.

Why this matters beyond Redmond

The tension here is fundamental to how AI gets deployed in corporate environments. Companies want access to the most powerful models available, and Fable 5 has been praised for its superior agent and coding capabilities. But capability means nothing if the legal framework around the model creates unacceptable exposure.

The fact that Microsoft historically maintained ZDR agreements with Anthropic makes the policy shift especially jarring. It suggests that Anthropic made a deliberate decision with its Mythos-class models to require data retention, possibly for safety monitoring, model improvement, or compliance with its own internal policies around advanced AI systems.

The split access policy—external customers approved, internal employees blocked—highlights how enterprise AI governance is becoming increasingly granular. Legal, compliance, and security teams now need to audit the entire data lifecycle for every model deployment, and those audits can produce different answers for different use cases within the same organization.