Updates to Fable 5 biology safeguards substantially reduce false positives and fallbacks, with testing showing about an 85% drop in biology-related fallbacks across product surfaces.

Internal testing confirms fewer fallbacks to Opus 5 when handling biology content, improving user experience for general biology inquiries.

Safeguards were rebuilt with input from internal and external experts and are acknowledged to still produce some false positives and not be perfect.

The article, dated August 7, 2026, is authored by Ethan Collins and reflects a contemporary update in AI safety and biology capabilities.

The update aims to broaden legitimate access for healthcare and academic sectors, while keeping high-risk queries blocked, and positions Fable 5 for responsible deployment across Pro, Max, Team, and Enterprise tiers at a rate of ten dollars per million input tokens.

The classifier and training data were refined to reduce benign biology requests triggering the classifier, while maintaining blocks on dual-use and harmful content.

Anthropic emphasizes building trusted access pathways for frontier biology capabilities, balancing openness with safety.

Claude Fable 5, a Mythos-class frontier AI launched in June 2026, targets complex tasks like scientific research and coding, with a brief export-control interruption in June that was resolved by early July.

Safeguards rely on safety classifiers that route restricted requests to Opus 5, with ongoing iterative tuning to balance false positives/negatives and resist jailbreaks.

Despite improvements, dual-use topics such as virology, toxicology, and molecular design still trigger a fallback to Opus 5 and are not open for professional biology research or drug development.

The biology safeguards were rewritten to handle a wider range of biology requests while cutting back on unnecessary fallbacks.

The update broadens access to biology-related content and general health and education topics while preserving safeguards against dual-use and bioweapon-related research.