Anthropic cuts Fable 5 biology fallbacks by 85%

Anthropic has updated Claude Fable 5’s biology safeguards to reduce false positives, allowing the model to answer substantially more legitimate biology-related requests. The company says biology fallbacks—cases where a request is redirected to the less capable Opus 5—fell by about 85% across its product surfaces during testing.

The change should be most visible in everyday health and educational use cases, such as interpreting lab results, understanding symptoms, and studying biology. Anthropic also expects healthcare professionals to receive more help with clinical tasks.

A narrower safety boundary

Fable 5 still routes requests involving areas such as virology, toxicology, and molecular design to Opus 5 when Anthropic considers them dual-use. That means the model is not yet positioned for unrestricted professional biology research or drug development.

Anthropic says it launched Fable 5 with unusually broad biology restrictions because the model can outperform experts on some complex biological tasks and provide operational assistance on others. The same capabilities that could support medical research could also help malicious actors, including those seeking to develop biological weapons.

The revised system relies on smaller AI classifiers that identify safeguarded biology requests. Anthropic rewrote the classifiers’ rules to describe permitted, beneficial uses in greater detail, gathered feedback from internal and external experts, created new training data, and retrained the system. The company says the updated classifier continues to flag harmful and dual-use research while blocking fewer benign queries.

For AI builders, the update highlights a practical safety tradeoff: broad initial restrictions can speed deployment but frustrate legitimate users, while better classifiers can expand access without removing safeguards. Anthropic says trusted access pathways remain part of its longer-term plan for frontier biology capabilities.

Source: Anthropic News