Anthropic has retuned the safety classifier that screens biology-related questions on Claude Fable 5, its flagship model, cutting the share of ordinary biology queries wrongly rerouted to a weaker fallback model by about 85%, the company said on August 7.
What changed
The classifier had erred toward caution, sending many everyday questions — interpreting lab results, understanding symptoms, or studying biology for a class — to Opus 5, a less capable model, whenever a query merely touched on biology. According to Anthropic, engineers rewrote the classifier’s “constitution,” the rule set defining what gets flagged, gathered feedback from internal and external experts, generated new training data, and retrained the system to hold the line on genuinely risky requests while releasing everyday ones.
Anthropic reported fallback-rate drops of roughly 67% on Claude.ai, 55% on the Cowork product, 17% on Claude Code, and 7% on the Claude developer platform. Healthcare professionals should also see fewer interruptions on routine clinical tasks, the company said.
What’s still blocked
The update narrows false positives rather than loosening the underlying policy. Fable 5 continues to refuse requests tied to dual-use professional research, virology, toxicology, and molecular design — categories Anthropic treats as carrying real biosecurity risk. The company has said biology and medicine represent one of AI’s biggest potential upsides, but that the same capabilities create a dual-use problem requiring careful gatekeeping rather than a blanket ban on the topic.
Why now
The change lands amid broader scrutiny of AI safety classifiers, which labs have been criticized both for over-blocking harmless queries and for gaps that let harmful ones through. The rewrite leans on Anthropic’s own Constitutional AI method, which trains and steers classifiers against a written rule set rather than relying on human feedback alone.
Anthropic did not give a timeline for extending the retuned classifier to other models in the Claude lineup.