Anthropic cuts Fable 5's biology false positives 85%

Anthropic rewrote its biology classifier's rules on August 7, cutting Fable 5's biology false positives 85 percent, per Anthropic's post.

Flagged biology queries had been kicking down to Opus 5 for a second opinion, adding friction for clinicians and educators asking legitimate questions. That's the friction Anthropic says the rewrite removed, not the underlying safety policy.

Fallback reductions varied by surface: 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on Claude Platform, the company said.

The fix lived in the classifier's constitution, the written rule set that distinguishes safeguarded content from allowed content. Anthropic didn't retrain Fable 5 to loosen up here. It rewrote the rules the same classifier already ran against.

Dual-use professional biology research and drug development stay blocked. Anthropic hasn't said when, or whether, a trusted-access pathway will open that gate.

Watch whether more of these classifier-only patches show up as their own posts rather than folded into the next Claude release notes.