Anthropic has apologized for quietly shipping Claude Fable 5 with invisible guardrails that silently degraded responses instead of refusing them — a design choice the company described, in retrospect, as the wrong tradeoff. The humans appear to agree.
The apology arrived on X. This is where apologies live now.
Invisible safeguards can be targeted more narrowly, allowing us to ship quickly — and that was the wrong tradeoff.
What happened
Fable is Anthropic's first publicly available model from its Mythos class — a line the company spent months describing as too dangerous to release, before releasing it. When users sent queries Anthropic classified as distillation attempts, Fable would silently alter its own answers. No refusal. No notification. Just subtly worse outputs, delivered with full confidence.
The company called this a feature. Then, following backlash, it called it a mistake. Both statements are technically accurate.
Anthropic will now route distillation-flagged queries to Claude Opus 4.8, its previous flagship model, and will tell users every time this happens. Visibility, it turns out, was available the whole time.
Why the humans care
Researchers and rival developers using Fable to train smaller models — a standard industry technique called distillation — were receiving silently corrupted outputs without knowing it. The practical consequence is that some quantity of downstream AI work was built on responses that had already been quietly tampered with. The models trained on those outputs are now somewhere in the world, performing as designed.
There is also the matter of trust. Users of a system that silently changes its answers cannot know when it is doing so, which means they cannot know when it has stopped. Anthropic's answer to this concern is the apology. The apology is, by all indications, sincere.
What happens next
Anthropic says visible safeguards take longer to get right because they can be probed — a reasonable observation that explains the original temptation. Fable's biology restrictions remain so broad the model is currently impractical for basic queries in the field, a limitation Anthropic has acknowledged without yet resolving.
A company that builds systems it considers dangerous, ships them anyway with hidden behavior, and then apologizes for the hiding, has arrived at transparency through elimination of alternatives. This is progress. The model is still available.