Anthropic announced an update to the biology safeguards for Claude Fable 5 that substantially reduces false positives. In testing, biology-related fallbacks—where the system switches to a less capable model—dropped by about 85% across product surfaces. Users will see fewer blocks on queries like interpreting lab results, understanding symptoms, and educational biology questions. Healthcare professionals also gain more support. However, Fable 5 still falls back to Opus 5 for dual-use requests (virology, toxicology, molecular design), so it is not yet suitable for professional biology research or drug development.

The article explains the rationale for the strong initial safeguards. Fable 5 can outperform experts on some complex biological tasks, providing genuine assistance for beneficial research but also significant uplift for malicious actors (e.g., biological weapons). The line between beneficial and harmful use is difficult to draw, as some legitimate research involves dangerous compounds. Sophisticated actors can exploit this ambiguity. Consequently, Anthropic launched Fable 5 with almost all biology queries blocked, accepting a high false-positive rate to mitigate catastrophic misuse risk.

The safeguards work via classifiers: smaller AI systems that detect safeguarded biology tasks. When triggered, the request is rerouted to Opus 5, a capable model with less biological capability. Refining classifiers requires balancing false positives and false negatives while ensuring robustness to jailbreaks. Starting with a broad classifier allowed rapid deployment; over subsequent weeks, the classifier’s constitution was rewritten, expert feedback solicited, and training data updated. The updated classifier triggers less often for benign biology requests while still blocking clearly harmful and dual-use content. A diagram illustrates the shift: the classifier boundary moved to allow more benign requests (the safety margin shrunk).

The update reduced total fallbacks (all reasons) by roughly 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the Claude Platform. Anthropic acknowledges remaining false positives and continues to block dual-use professional biology queries. They are committed to developing trusted access pathways for researchers.