Anthropic Cuts Fable 5 Biology Refusals 85%

Anthropic rewrote the rule set governing Fable 5's biology safety classifier, cutting biology-related fallback routing by about 85%, per the company's August 7 post.

At launch, that classifier routed flagged biology queries down to the weaker Opus 5 model, producing refusals on legitimate clinical and educational questions. For researchers, students, and clinicians using Claude, that meant losing access to Anthropic's strongest model over queries that were never actually dangerous.

Total fallback reductions ran 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on Claude Platform, according to Anthropic.

Dual-use professional biology research and drug development queries stay blocked. Anthropic says it's building trusted-access pathways for those cases, but hasn't shipped them yet. Until it does, the loosened classifier still draws a hard line at anything resembling weapons-relevant biology work.

The take

An 85% cut this specific is an admission. The original classifier was tuned so conservatively it blocked real users more often than real threats. Watch whether Anthropic publishes a similar rewrite for other Fable 5 safety categories. Biology may have just been the category drawing the most complaints.

A related report found agent containment shipping across three unrelated products in 48 hours. Another logged Codex tightening its automatic-review defaults for cyber-capable models. Both cover different problems, agent behavior and code review rather than biology content filtering. But the pattern rhymes: ship a conservative filter first, loosen it once real users complain.