We’re updating Claude Fable 5’s biology safeguards to reduce false positives. In

Anthropic has retrained the filter that decides which biology questions its frontier model, Claude Fable 5, is allowed to answer. The filter is a safety classifier, a smaller automated model that watches for safeguarded biology requests or harmful outputs. When it fires, the request is rerouted to Opus 5, which Anthropic describes as capable but with less biological capability than Fable 5. That reroute is the fallback users see. In Anthropic's account, the update cut biology-related fallbacks by about 85% across its products in testing, so questions such as interpreting lab results, understanding symptoms or learning biology should fall back far less often.

The original classifier was deliberately broad. Anthropic says it launched Fable 5 with almost all biology queries blocked, accepting many false positives, so the model could ship for other uses while the classifier was refined, rather than holding it back for weeks or months. The fix came from rewriting the classifier's constitution, the set of rules that separates safeguarded content from allowed content, with benign uses carved out in detail. Anthropic says it gathered feedback from internal and external experts, built new training data from those rules, retrained the classifier and checked that it still generally fires on harmful and dual-use research content. It expects total fallbacks for any reason to fall by roughly 67% on Claude.ai, 55% on Cowork, 17% on Claude Code and 7% on the Claude Platform.

Professional research biology is still blocked on purpose. Requests Anthropic considers dual-use, meaning they could serve beneficial or harmful purposes, including virology, toxicology and molecular design, continue to go to Opus 5, so Anthropic says Fable is not yet usable for professional biology research or drug development. Its reason is that its own capability assessments show Fable 5 could give significant uplift to someone trying to develop a biological weapon, and that legitimate and harmful work can look alike: a live vaccine requires growing the pathogen it prevents. Anthropic says it plans to close that gap through trusted access pathways for researchers, and acknowledges that some low-risk requests will still be caught.