The Verge reports that Microsoft has pulled Claude Fable 5 from the model picker its own employees use inside internal GitHub Copilot, after Anthropic introduced new data-retention rules for the model. The objection is narrow and concrete. Fable 5 mandates a 30-day retention of prompts and outputs, with items that Anthropic's classifiers flag kept for as long as two years, and, crucially, it cannot operate under a Zero Data Retention agreement. Every other Claude model on the API can, Opus 4.8, Sonnet 4.6, and Haiku 4.5 all support ZDR, and those remain available to Microsoft employees internally. Fable 5 is the one that got pulled, and Microsoft has told staff its legal teams are evaluating the retention change over concerns about customer data and confidential information.
This is the enterprise-compliance bill arriving for a safety design we covered on launch, and the two are the same mechanism. When Fable 5 shipped, the headline was graceful capability degradation: classifiers route risky cyber, bio, and distillation requests down to Opus 4.8 instead of refusing. The 30-day retention was the fine print underneath, and it is not incidental. To make that safety case auditable, to show that the routing did what it claims and to investigate misuse, the system has to keep the traffic, which is precisely why Zero Data Retention is not on the menu for this model and is for the others. The safety mechanism and the retention mandate are not two policies that happen to coincide; they are one policy seen from two sides.
It connects directly to yesterday's complaint from a different direction. Security researchers objected that Fable 5's classifier over-blocks legitimate defensive work; Microsoft's objection is that the same classifier's retention requirement is itself a blocker for anyone with confidentiality obligations. Over-blocking and over-retaining are both downstream of the same design choice, to make the model's safety legible by inspecting and keeping what flows through it. For a regulated bank, a hospital, or any shop with strict data-handling contracts, mandatory retention is not a footnote, it is a reason the most capable public tier is simply unusable as shipped.
The wrinkle on top is the identity of the objector. Microsoft resells Claude through Azure and Copilot and is one of Anthropic's largest distribution partners, and it is the one restricting this model from its own engineers. Beyond the optics, it is a clean data point on the cost side of the access-stratification trend the frontier is moving toward: the newest and strongest public tier arrives with a data-handling string attached that some of the most demanding customers cannot accept, so capability and compliance pull apart rather than rising together. The practical note for builders in regulated environments is direct: before you route confidential work through a frontier model, check whether it offers Zero Data Retention, because the newest tier may be exactly the one that does not.