Anthropic published on 7 August that it has rewritten the biology classifier guarding Claude Fable 5, cutting biology-related fallbacks by about 85%. The company says Fable 5 shipped with deliberately very broad biology blocking to manage dual-use risk, and that the breadth produced excessive false positives.
Fallback, not refusal
The mechanism decides how to read the numbers. Fable 5 does not decline a flagged biology request; a classifier detects it and reroutes the query to Opus 5, described as "a capable model that does not have the same level of biological capability". The percentages describe how often requests get rerouted, not how often they were denied.
Uneven across products
Total fallback reductions land very differently by surface: 67% on Claude.ai, 55% on Cowork, 17% on Claude Code and 7% on Claude Platform. The headline 85% is biology-specific; the all-cause figures vary ninefold, which puts most of the recovered ground in consumer chat rather than developer traffic.
What is still blocked
Nothing was relaxed on the dual-use side. Professional biology and drug-development queries touching virology, toxicology and molecular design still fall back. The named improvements are interpreting lab results, understanding symptoms and learning biology in an educational context — the cases the broad filter was catching by accident.
The rare part is the disclosure
The change itself is unglamorous: carve out benign uses in detail, build training data, retrain. What is unusual is publishing the before and after — a frontier lab putting a number on how much legitimate work its most aggressive safeguard was blocking, and how much of that was recoverable without touching the model.