Microsoft has pulled Claude Fable 5 from the internal model picker its employees use inside GitHub Copilot, one day after Anthropic released the model on June 10, 2026. The block is driven by Anthropic's new data retention requirements, which keep prompts and outputs for 30 days and, in some cases, up to two years — terms that conflict with Microsoft's Zero Data Retention policy for employee tools.
Fable 5 is the first broad release in Anthropic's Mythos class, the same family the company said weeks ago was too capable at cybersecurity tasks to ship publicly without safeguards. Anthropic stripped those constraints by layering safety classifiers on top of the model, and those classifiers are what require the retention window in the first place.
“Microsoft is limiting the use of Claude Fable 5 for employees because of Anthropic's new data retention requirements”— Tom Warren, The Verge Senior Correspondent
Microsoft did not block the model from paying customers. Fable 5 went live for external GitHub Copilot and Foundry users on launch day, and remains available there. The restriction applies only to the version of Copilot that Microsoft's own employees use day to day.
Key facts
- 01Microsoft removed Claude Fable 5 from the internal GitHub Copilot model picker used by employees, one day after Anthropic released it on June 10, 2026.
- 02Anthropic retains prompts and outputs from Fable 5 for 30 days to run new safety classifiers, breaking Microsoft's Zero Data Retention policy.
- 03Prompts flagged as violating Anthropic's usage policy can be stored for up to two years.
- 04All other Claude models remain available to Microsoft employees because they still run under ZDR terms.
- 05Microsoft has continued offering Fable 5 to external GitHub Copilot and Foundry customers while its legal team reviews the change.
Every other Claude model — Opus, Sonnet, Haiku, and prior Fable releases — is still selectable in the internal picker. Those models operate under Anthropic's Zero Data Retention agreement with Microsoft, under which no prompts or outputs are retained on Anthropic's side. Fable 5 is the first Claude release that breaks that contract.
The retention mechanics are specific. Anthropic holds prompts and outputs for 30 days to feed its new safety classifiers, then deletes them. Anything the classifiers flag as a usage-policy violation can be stored for up to two years. For a company whose employees routinely paste customer data, unreleased product details, and internal code into Copilot, both windows are legally fraught.
“All other Claude models are still available internally at Microsoft, because they operate under Zero Data Retention (ZDR) rules”— Tom Warren, The Verge Senior Correspondent
Microsoft's legal teams are evaluating whether the new terms can be reconciled with the company's confidentiality obligations. According to people who spoke to The Verge, the main concerns center on customer data and confidential information, and there is no internal timeline yet for clearing or permanently rejecting Fable 5. Microsoft declined to comment.
The episode is a sharp illustration of how safety policy and enterprise procurement now collide directly. Anthropic added the retention window precisely because Mythos-class models can answer cybersecurity and biosecurity queries that earlier Claude versions refused — the classifiers exist to catch misuse in flight. But the cost of keeping that data, even briefly, is that some of the most lucrative enterprise customers cannot legally use the model on sensitive workloads.
It also lands awkwardly inside the Microsoft-Anthropic relationship. Microsoft only added Claude to Copilot and Foundry late last year as a hedge against single-vendor dependence on OpenAI, and Fable 5 was meant to be the headline addition for coding agents this summer. Shipping it to customers while blocking it for staff is the kind of split that procurement teams at other regulated enterprises — banks, hospitals, defense contractors — will notice immediately.
Anthropic has not publicly addressed how it plans to offer Mythos-class capabilities to ZDR customers, and there is no indication yet whether a separate enterprise tier with shorter retention or on-premise classifier inference is in the works. The trade-off Anthropic chose — safety classifiers in exchange for a retention window — is the one that's currently on the table, and Microsoft is the first major customer to publicly balk at it.
We covered the underlying classifier system last week when cybersecurity researchers reported that Fable's safeguards were blocking routine code-review prompts. The Microsoft restriction is the commercial echo of that same engineering choice: the more aggressive the on-server safety layer, the more data Anthropic has to keep, and the harder it becomes to sell the model into environments where keeping that data is itself the problem.
For Anthropic, the question now is whether the Mythos retention model holds across the enterprise base or whether Microsoft's pause becomes a template. ZDR is not a Microsoft-only concern — it is the default ask from every Fortune 500 legal team evaluating frontier models, and the GitHub Copilot internal picker is one of the most visible deployments of Claude anywhere. If Microsoft's lawyers don't clear Fable 5, Anthropic will be selling its flagship Mythos release into a market where its largest distribution partner won't run it in-house, which is a difficult position to defend through the rest of the year.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
