Anthropic released Claude Fable 5 on Tuesday. It’s the first model in their new Mythos class, the same family Anthropic said a few weeks ago was so good at hacking it was too dangerous to release. They bolted on safeguards and shipped it anyway.
Microsoft moved fast. Within hours, Fable 5 was live for GitHub Copilot and Foundry customers. The people paying Microsoft for AI got the newest, sharpest model the same day it dropped.
Microsoft’s own employees did not.
The Verge
reported that Fable 5 is missing from the internal model picker Microsoft staff use for their version of Copilot. Every other Claude model is still there. Just not the new one. Microsoft’s legal team is reviewing it and hasn’t cleared it for internal use.
Sit with that. Microsoft will happily sell you a tool it won’t let its own engineers run on company laptops.
The reason isn’t that Fable 5 is bad. It’s a setting called data retention, and once you understand it, you’ll start checking it on every AI tool you touch.
Here’s the mechanics.
Older Claude models at Microsoft run under something called Zero Data Retention, or ZDR. You send a prompt, the model answers, and nothing gets stored. The conversation evaporates. That’s why legal was fine with it.
Fable 5 breaks that. To run Anthropic’s new safety classifiers, the ones that keep the model from helping someone build malware, Anthropic has to
keep your prompts and the model’s answers. They hold that data for 30 days by default. If something you typed trips a policy flag, they can hold it for up to two years.
So if a Microsoft employee pasted confidential customer data or unreleased product plans into Fable 5, that text would sit on Anthropic’s servers for a month. Maybe longer. For a company that lives and dies by trade secrets, that’s a hard no.
Microsoft declined to comment, which is corporate for “yeah, that’s exactly what’s happening.”asked for seconds.