Anthropic retunes Claude Fable 5 to cut biology false alarms by 85%

Anthropic said Aug. 7 it rewrote and retrained the biology-safety classifier behind Claude Fable 5, cutting biology-related false-positive fallbacks by about 85% in testing while still blocking higher-risk dual-use requests.

The retuned classifier – Anthropic calls its rule set a “constitution” – carves out benign-use exceptions so Fable 5 can answer more everyday health and educational questions, such as interpreting lab results, the company said.

By product surface, the change cut total fallback volume by roughly 67% on Claude.ai, 55% on Cowork, 17% on Claude Code and 7% on the Claude Platform.

Fable 5 still hands off to the more capable Claude Opus 5 for requests Anthropic flags as higher-risk, including virology, toxicology and molecular design. The company said it is building “trusted access pathways” for qualified researchers who need those capabilities.

Fable 5 launched June 9 with deliberately conservative safeguards and, by Anthropic’s own description, remains “not yet usable for professional biology research and drug development.”

The 85% figure is Anthropic’s own testing result, not an independent audit. Loosening a safety filter to reduce false positives inevitably shifts where the line sits between over-blocking benign queries and letting a harmful one through, and the company has not published the tests behind the number.