Anthropic loosens Fable 5's biology filter to cut false blocks

Safety filters that block harmful biology questions have a well-known problem: they also block harmless ones. A patient asking about lab results or symptoms can trip the same rules meant to stop bioweapon design. Anthropic says it rewrote the classifier rules behind Fable 5 to draw that line more carefully, retraining the system on updated data and taking feedback from internal and external experts before shipping the change on August 7.

The effect is large and measured per product. Biology-related fallbacks, cases where the model refuses or hedges, dropped by about 85% overall. The reduction was 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the Claude Platform, which suggests the over-blocking was worst in the consumer chat product and much smaller for developers. Everyday questions like interpreting a blood test or understanding a symptom should now go through more often. The harder guards stay in place: Fable 5 still refuses genuine dual-use requests in virology, toxicology, and molecular design, and it routes frontier biology work to trusted access pathways rather than answering openly.

This is a tuning problem more than a policy reversal, and the honest read is that the earlier filter was miscalibrated toward refusal.

Why it matters

If you are a clinician, student, or researcher who kept hitting refusals on ordinary biology and health questions, this directly targets that friction. It also shows how labs are recalibrating safety filters after over-blocking, which is worth watching if you build on these models and need to predict when they will say no.