Anthropic retunes Claude’s biology safeguards, cutting blocked queries by 85%

10:00 AM PT · August 11, 2026

Anthropic announced updates to the safety classifiers that govern how Claude Fable 5 handles biology related questions, aimed at reducing unnecessary refusals without loosening protections against genuinely dangerous, dual use requests such as those tied to bioweapons. The company says the retuned classifiers cut biology related fallbacks, cases where Claude declines or hedges on a legitimate question because it resembles a risky one, by roughly 85% across its product surfaces, with the improvement breaking down unevenly: about 67% fewer fallbacks on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the Claude Platform for developers. Anthropic frames the change around real use cases that were getting caught by the old, more conservative classifier, such as people trying to interpret their own lab results, understand symptoms, or learn about biology in an educational setting, as well as healthcare professionals doing clinical work that touches on sensitive topics like pathogens or toxins. The company says the underlying dual use safeguards, meant to catch requests that could meaningfully help someone create a biological weapon, remain unchanged in strength; what changed is the classifier’s ability to tell those apart from ordinary biology questions. The update is part of a broader pattern of Anthropic publishing specific, numeric detail about how its safety systems perform, rather than describing changes only in general terms.