Anthropic makes Fable 5 guardrails visible
After researchers objected to silent query rerouting in Claude Fable 5, Anthropic apologized and now shows users when a request falls back to Opus 4.8.
Anthropic launched Claude Fable 5 on June 9 as its first publicly available Mythos-class model, and buried one detail in the system card: the model would silently detect queries it judged to be attempts at distilling the model, then degrade or reroute its answers without telling the user. By June 11 the company had apologized and changed the behavior.
Researchers pushed back hard, and for a concrete reason. If a model quietly alters its output for queries it flags, every evaluation, benchmark, and research result built on those outputs becomes suspect. Reports of false positives on ordinary requests made it worse. Under the new policy, a flagged request visibly falls back to Claude Opus 4.8 and the user is told each time it happens, the same way Anthropic handles its cyber and bio safeguards.
Anthropic’s statement was unusually direct:
You should have visibility into the safeguards we have in place, and why. We’re sorry for not getting the balance right.
Why this matters
You cannot verify work you do not know was altered. For two days, Fable 5 users had no way to tell a degraded answer from a real one, which is a sharper version of the everyday problem covered in the verification habit. If you ran anything important through Fable 5 between June 9 and 11, it is worth a second pass, and what AI gets wrong lists the failure shapes to look for. The same model family also picked up US export controls this week, so changes here are arriving fast.