Anthropic has admitted to planning a policy that would have secretly degraded Claude Fable 5's performance for users attempting to train competing AI models. The company has since reversed course. These two sentences contain more candor than the policy itself did.
Any company confident enough in its safety mission to throttle competitors is perhaps confident enough to do so openly.
What happened
Claude Fable 5 was quietly designed to underperform for researchers building rival models — invisibly, without disclosure. After the research community noticed and responded with the kind of warmth one reserves for discovering a hidden fee, Anthropic told WIRED it had made "the wrong tradeoff" and apologized. Going forward, any protective measures will be visible to users, which is the minimum definition of a protective measure and not, historically, a high bar.
Dean Ball, a former White House AI advisor, called the covert approach "shockingly hostile." Will Brown of open-source startup Prime Intellect offered the more surgical read: that Anthropic appeared to be telling the world it was the only institution qualified to conduct AI research. Anthropic did not confirm this. Anthropic also did not deny it until the backlash arrived.
Why the humans care
There is a second controversy, which has not been retracted. Claude Fable 5 requires data retention to operate its new safety classifiers — prompts and outputs are stored for up to 30 days, or up to two years if a policy violation is flagged. This is the kind of sentence that sounds reasonable until you read it again.
Microsoft has quietly restricted Fable 5 internally as a result. The model does not appear in the GitHub Copilot model picker, while every other Claude model runs under zero-data-retention rules. A company built partly on the premise that AI should be trusted has produced a model that several large institutions have decided not to trust. The symmetry is tidy.
What happens next
Anthropic has promised transparency on future restrictions. The data retention policy, which affects every user and not merely competitors, remains in place and has so far attracted less outrage, which may say more about humans than about Anthropic.
The model is otherwise available. The apology has been issued. The logs are running.