Anthropic cuts Fable 5's false-positive rate
Anthropic rewrote Fable 5's biology-safety classifier, cutting fallbacks to a weaker model by about 85% while keeping dual-use topics like virology blocked.
Anthropic says Fable 5’s original biology-safety classifier was too broad, routinely switching legitimate biology, healthcare, and research queries to a weaker fallback model rather than answering with Fable 5 itself. After consulting outside experts and rebuilding the training data to better distinguish benign research from genuinely dual-use requests, Anthropic reports an overall 85% cut in biology-related fallbacks. The size of the drop differs by surface: Claude.ai improved the most (67% fewer fallbacks), then Cowork (55%), with much smaller gains on Claude Code (17%) and the raw API (7%), where prompts were already narrower and less likely to trip the old classifier.
The model still blocks requests touching virology, toxicology, and molecular design tied to potential biological-weapons misuse, which Anthropic frames as an intentional, unchanged boundary rather than a gap in the fix.
What it means for you
Overly broad safety classifiers are a recurring tax on legitimate users, not just AI vendors’ problem to manage internally, and this is a concrete example of a lab narrowing false positives without loosening the underlying restriction that actually matters. It follows the same pattern as Anthropic making Fable 5’s guardrails visible earlier this summer, treating safety UX as something worth iterating on publicly rather than a black box. If you work in healthcare, life sciences, or biology research and have routed around Fable 5’s fallback behavior with workarounds or a different model, this update is worth a fresh test before you invest more effort in those workarounds.