Improving Fable 5's biology safeguards

Updates have been made to Claude Fable 5's biology safeguards to substantially reduce false positives and system fallbacks. Users will now experience roughly an 85 percent reduction in biology-related fallbacks across product surfaces, meaning the system will switch to a less capable model much less frequently when handling everyday health, educational, and clinical tasks.

When Fable 5 originally launched, it utilized very broad safety classifiers that blocked nearly all biology queries out of an abundance of caution, preventing potential misuse in dual-use domains like virology, toxicology, and molecular design. Because frontier AI models like Fable 5 can outperform experts on complex biological tasks and potentially provide operational support for creating biological threats, the initial broad restrictions were put in place to prevent catastrophic misuse while managing the difficulty of distinguishing between beneficial and harmful research.

To achieve the new updates, the classifier's constitution was rewritten with input from internal and external experts, and the training data was retrained to better distinguish between benign and dual-use content. Despite these improvements, the system will still block dual-use professional biology and drug development queries, and low-risk requests within the safety margin may still occasionally trigger fallbacks as development continues.