Anthropic’s latest model drop, Claude Fable 5—the first in its Mythos-class—has already hit a compliance wall inside Microsoft. According to internal sources, the company has restricted employee access to the model in its internal GitHub Copilot tooling, citing unresolved legal concerns around Anthropic’s revised data retention requirements.

While Microsoft moved quickly to offer Fable 5 to external customers via GitHub Copilot and Azure Foundry, the model is conspicuously absent from the internal model picker used by employees. All other Claude models remain available under the existing Zero Data Retention (ZDR) agreement, which ensures prompts and outputs are not stored. Fable 5 breaks that pattern.

The core issue: Anthropic requires data retention to power new safety classifiers embedded in Fable 5. Prompts and outputs are stored for 30 days by default; flagged violations can be held for up to two years. For Microsoft’s legal team, that raises red flags around customer data and confidential internal information flowing through a third-party retention pipeline.

This is a classic engineering trade-off between safety and privacy. Anthropic designed Fable 5 with guardrails tight enough that it deemed the broader Mythos class too dangerous to release publicly just weeks ago. Those same guardrails now require persistent storage of user interactions—a non-starter for many enterprise deployments.

Microsoft’s legal review is ongoing. Until it clears the model for internal use, employees are stuck with older Claude versions that don’t log their work. Anthropic declined to comment. Microsoft did not respond by deadline.

Source: The Verge