Microsoft restricts engineers' access to Anthropic's Claude Fable 5 over data retention
L'essentiel
- Microsoft has blocked its own engineers from accessing Anthropic's new Claude Fable 5 AI model due to its data retention policies, while still offering it to customers.
- The model's 30-day data hold conflicts with Microsoft's Zero Data Retention (ZDR) rules for internal use.
Résumé généré par IA
Pourquoi c'est important
Microsoft has restricted its own engineers' access to Anthropic's new Claude Fable 5 AI model due to its data retention requirements, which differ from the Zero Data Retention (ZDR) policies applied to other Claude models. This restriction applies internally, while customers can still access the model.
Anthropic launched Claude Fable 5 on Tuesday, its first Mythos-class model and the most powerful one the company has cleared for wider use. Within a day, Microsoft had quietly walked it out of the building. The Verge reports that Microsoft is restricting employee access to Fable 5 over Anthropic's new data retention requirements, even as it rolled the same model out to its GitHub Copilot and Foundry customers without missing a beat. Every other Claude model is still available to staff, because they run under Zero Data Retention. Fable 5 does not, and that single technical difference has Microsoft's legal team holding the line. The split treatment is unusual, especially given how deeply the two companies are now wired together, from the November Foundry deal covering Sonnet 4.5, Opus 4.1, and Haiku 4.5 to the Claude tech now sitting inside Microsoft 365 Copilot. It lands at an awkward moment in a partnership that has otherwise been getting closer, not more cautious.
The model picker has a Fable-shaped hole in it
The internal version of GitHub Copilot that Microsoft staff actually use no longer shows Claude Fable 5 in the model dropdown, according to The Verge. Every other Claude option is still sitting there, untouched. Microsoft has told staff that legal is still evaluating Anthropic's changes, and that it's not yet clear whether Fable 5 will be approved for internal work at all.
Why a 30-day clock is suddenly a legal headache
Claude Fable 5 needs data retention to run Anthropic's new safety classifiers. Prompts and outputs are held for 30 days, and anything flagged as a usage-policy violation can sit on Anthropic's servers for up to two years. For a company that handles customer data and confidential workstreams at Microsoft's scale, those two windows are big enough to make in-house counsel want a longer look. Anthropic's own documentation confirms Mythos-class models are not available under ZDR on the Claude API.
Anthropic Claude Fable 5 data retention vs Microsoft's ZDR rules
The contrast with Anthropic's other models is what makes this awkward. ZDR-eligible Claude versions return nothing to Anthropic's storage after the API response lands. Fable 5 is a Covered Model, which means retention is the price of entry. On Microsoft Foundry, however, retention rules are set by Microsoft itself rather than Anthropic, which is part of why the customer-facing rollout could go ahead while the internal one stalled. Microsoft's own engineers, it turns out, are held to a stricter standard than its paying customers. Fable 5 arrives weeks after Anthropic said the Mythos family was too capable at cybersecurity tasks to ship publicly. The company added prompt-level safeguards, and Fable 5 is the toned-down version that made it through. The retention requirement is part of that safety net, since the classifiers need something to read. Microsoft didn't respond to The Verge in time for publication, and for now, the company's own engineers stay locked out of a model their customers can already call.
Questions ouvertes
- Will Microsoft eventually approve Fable 5 for internal use?
- What specific legal concerns are driving Microsoft's decision?
- How will this impact the broader Microsoft-Anthropic partnership?
- Will other companies face similar issues with Fable 5 or future models?
