Claude Fable reportedly flags benign document work as biology
Gergely Orosz said Claude downgraded benign Fable requests to Opus and capped output after classifying the work as risky. He traced one [bio] flag to a Google Sheet collecting public social-post replies.
TL;DR
- Fable reportedly routed a request to organize social-post replies to Opus 5 after returning a
[bio]safeguard label, according to GergelyOrosz's follow-up and the attached error panel. - The input GergelyOrosz later isolated was a Google Sheet collecting replies to two public posts on X and LinkedIn, as his later post described it.
- The Fable interface says deliberately broad safeguards can flag legitimate coding, cybersecurity, and biology work, a warning shown in the error panel.
- The model switch was visible in the reported failure, while Anthropic's launch announcement says Fable 5.1 and restricted-access Mythos 5.1 share a model but use different safeguard levels.
Anthropic's announcement claims newer cybersecurity safeguards cut false positives by 60 percent. This report carries a terse [bio] bucket instead, and the identified input was a spreadsheet of public replies rather than a life-sciences artifact. One reply says the shared request will be reviewed.
The `[bio]` route
GergelyOrosz said the rejected job was to go through responses to a thread and group them. In a follow-up, he described that request as the one that triggered the system's alarm.
He later wrote in another post that Fable imposed word limitations, despite his claim that the work involved no cybersecurity, biology, offensive, or defensive activity.
Safeguard levels
The interface frames the fallback as the cost of broad controls that can occasionally catch legitimate tasks, according to the displayed message.
Anthropic's launch announcement describes Fable 5.1 and Mythos 5.1 as the same model at different safeguard levels. Fable is generally available; Mythos is reserved for trusted-access programs supporting cybersecurity and life-sciences work. Its 60-percent false-positive figure is specifically for cybersecurity safeguards, not an explanation of this reported [bio] classification.
The public-reply spreadsheet
Four days after the initial report, GergelyOrosz said he had traced the trigger to a Google Sheet gathering responses from two public X and LinkedIn posts. He speculated that Fable may have failed to interpret the phrase “death of the UI.”
After GergelyOrosz shared the attempted request, Addy Osmani's reply said, “Will take a look.”