Anthropic shipped Claude Fable 5 yesterday, its first Mythos-class model, and the release is already causing friction at one of its biggest partners. According to The Verge AI, Microsoft is restricting employee access to Claude Fable 5 because of Anthropic’s new data retention requirements. The model went out fast to GitHub Copilot and Foundry customers, but The Verge AI reports it’s missing from the internal model picker that Microsoft’s own staff use for company versions of Copilot.
What stands out here is the split. Every other Claude model is still available to Microsoft employees, because those run under Zero Data Retention rules. Fable 5 breaks that pattern, and Microsoft’s legal teams are now deciding whether it’s safe to let staff use it at all.
What’s actually changed
The sticking point is how Fable 5 handles your prompts. To run Anthropic’s new safety classifiers, the model needs to retain data. Here’s what that means in practice, per The Verge AI:
- Anthropic keeps prompts and outputs, then deletes them after 30 days.
- Anything flagged as violating Anthropic’s usage policy can be stored for up to two years.
- All previous Claude models avoided this entirely under Zero Data Retention.
For a company like Microsoft, that’s not a small detail. The concern, sources told The Verge AI, is customer data and confidential information passing through a model that holds onto prompts. Microsoft has reportedly been telling employees its legal teams are evaluating the changes. It’s not yet clear they’ll clear Fable for internal use.
Microsoft declined to comment.
Why this matters
This is significant because it shows the cost of safety landing on the customer. Anthropic built Fable 5 with prompt safeguards specifically because the Mythos class is so capable. Just weeks ago, the company said the family was so strong at cybersecurity tasks that it was too dangerous to release publicly. The data retention requirement is the trade-off that made a public release possible. Those classifiers need to see and store prompts to do their job.
So Microsoft is caught in an awkward spot. It’s selling Fable 5 to its cloud customers through Foundry and Copilot while simultaneously keeping it away from its own engineers. That tells you the retention terms aren’t a minor compliance footnote. They’re enough to make a sophisticated buyer hit pause internally even as it resells the same model.
The broader signal: as frontier models get more capable, the safety machinery around them gets heavier. Zero Data Retention has been a baseline expectation for enterprise AI buyers who handle sensitive data. Fable 5 asks those buyers to give that up. Microsoft’s reaction is the first public sign that the ask is a real friction point, not a rubber stamp.
What to watch next
A few things worth tracking if you work with these models:
- Whether Microsoft clears Fable internally. Its legal review is the bellwether. If Microsoft’s own lawyers won’t sign off, other large enterprises will ask the same questions.
- How other regulated buyers respond. Banks, healthcare firms, and government contractors lean hard on data retention guarantees. Expect them to scrutinize the 30-day default and the two-year flag window.
- Whether Anthropic offers a carve-out. The obvious resolution is a Zero Data Retention tier for Fable, but that may not be technically possible if the safety classifiers require retained prompts to function.
For practitioners, the practical takeaway is simple. Before you route confidential or customer data through Fable 5, check your own retention terms and what your legal team will tolerate. The capability jump is real, but so is the data handling change that comes with it. The model that’s powerful enough to need new safeguards is also the model that asks you to store more than you used to.
Microsoft’s internal pause may get lifted next week or it may harden into policy. Either way, it’s an early read on how the industry will weigh capability against control. You can find the full reporting at the original source.