Anthropic cuts Claude Fable 5 false blocks by 85%

Anthropic has rolled out an update to the biology safeguards of its Claude Fable 5 model, drastically lowering the frequency of false positives. Previously, the system would frequently trigger a fallback mechanism, routing biology-related queries to the less capable Opus 5 model to prevent potential misuse. With the new update, these biology-specific fallbacks have dropped by approximately 85 percent across Anthropic's product surfaces. This change translates to a reduction in overall fallbacks of roughly 67 percent on Claude.ai, 55 percent on Cowork, 17 percent on Claude Code, and 7 percent on the Claude Platform.

To achieve this, Anthropic rewrote the safety classifier's constitution, which defines the rules separating benign queries from dangerous ones. The company then retrained the classifier on updated data and verified its performance. These strict guardrails are necessary because Fable 5 can outperform human experts on certain complex biological tasks. In the wrong hands, such capabilities pose severe dual-use risks, such as assisting in the creation of biological weapons. The US Intelligence Community's 2026 Annual Threat Assessment highlighted that biotechnology advances could lead to novel biological threats, prompting Anthropic to initially launch Fable 5 with highly restrictive blocks.

For practitioners, this update significantly expands the practical utility of Fable 5 in everyday scenarios. Healthcare professionals and educators can now use the model for clinical tasks, interpreting lab results, understanding symptoms, and teaching biology without constantly triggering safety blocks. However, because the model still routes dual-use queries to Opus 5, it remains restricted for professional biology research and drug development. Tasks involving virology, toxicology, and molecular design will still trigger the fallback mechanism until Anthropic establishes trusted access pathways for these frontier capabilities.