Improving Fable 5's biology safeguards

| Source: Anthropic News (community RSS)

Tags: Anthropic, Fable 5, biology safety, AI safety, biosecurity, healthcare AI

Anthropic updated Fable 5's biology safety filters, cutting false positives by ~85% — everyday health questions, lab result interpretation, and educational biology will no longer trigger fallback to a less capable model, though dual-use requests (virology, toxicology) still route to Opus 5.

Details

Anthropic made targeted improvements to Fable 5's biology classifier, reducing the rate of "fallbacks" — where the system switches to a less capable model after detecting a potentially sensitive biology query — by approximately 85% across its product surfaces. In practice, users will stop hitting model downgrades on routine tasks: interpreting lab results, understanding symptoms, general biology education. Healthcare professionals gain more consistent access to Fable 5 capabilities on clinical questions. The safety perimeter wasn't removed — it was narrowed. Fable 5 still routes to Opus 5 for requests Anthropic considers dual-use: virology, toxicology, and molecular design remain restricted. Anthropic's public rationale is explicit: Fable 5 can outperform experts on some highly complex biological tasks, and the company believes it could provide significant uplift to a malicious actor attempting bioweapons development. The announcement previews a planned 'trusted access pathway' for professional biology researchers who currently can't use Fable 5 for drug development or frontier biology — acknowledging the current safety calibration is too broad for legitimate expert use cases.