Anthropic eases Fable 5 biology blocks: 85% fewer false refusals
A retrained classifier stops walling off everyday health, lab-result and biology-learning questions — biggest relief on Claude.ai and Cowork, not the API.
Copy markdown
Biology false-positives cut ~85%
Anthropic rewrote Fable 5's biology classifier — retraining on updated data with expert feedback — and cut biology-related fallbacks by about 85% across its product surfaces. Protections on genuinely harmful content stay; what shrank is the cautious margin that was blocking prompts flagged “very likely benign” and stopping sessions mid-task.
Where you'll actually feel it
Total fallbacks drop 67% on Claude.ai and 55% in Cowork, but only 17% in Claude Code and 7% on the Claude Platform API. If you use Claude for research, health, or bio-adjacent writing, that means far fewer dead ends; if you build against the API, expect little change, since that traffic rarely tripped the biology filter.
What's back in scope
Everyday health and education are unblocked: interpreting lab results, understanding symptoms, and learning biology concepts, plus better support for healthcare pros on clinical tasks. Benign prompts that used to bounce — down to “how babies are made” — should now go through.
What still gets walled off
Dual-use stays gated: virology, toxicology, and molecular design — plus professional biology research and drug development — still route to a restricted Opus 5 fallback rather than a full Fable 5 answer. Anthropic says vetted researchers will get real access via a separate “trusted access” pathway, not prompt-wrangling.