Anthropic has updated the biology safeguards on Claude Fable 5, cutting false-positive "fallbacks" by around 85% across its product surfaces, the company announced on 7 August 2026.
Fallbacks occur when a safety classifier detects a biology-related query and reroutes it to Opus 5, a less capable model, rather than letting Fable 5 respond directly. Anthropic said the change means fewer everyday health and educational questions, such as interpreting lab results or understanding symptoms, will be needlessly blocked, and that healthcare professionals should see more support on clinical tasks.
The company said Fable 5 still falls back to Opus 5 for dual-use domains including virology, toxicology and molecular design, meaning it remains unsuitable for professional biology research or drug development. Anthropic said distinguishing beneficial from harmful biology use is often genuinely difficult, since legitimate research sometimes requires producing the same dangerous substances it aims to guard against, as with live vaccines grown from the pathogens they prevent, or the hypertension drug captopril, developed from toxic compounds isolated from snake venom.
Anthropic said its capability assessments show Fable 5 could offer significant uplift to a malicious actor attempting to develop a biological weapon, and pointed to the US Intelligence Community's 2026 Annual Threat Assessment, which it said states that several state actors likely maintain active offensive biological and chemical weapons programmes that could be accelerated by frontier AI.
The update followed several weeks of rewriting the classifier's underlying rules, developing new training data, and testing with internal and external experts to distinguish dual-use research content from benign queries. Anthropic said it remains committed to building trusted-access pathways that would eventually let vetted biology researchers use its most capable models.
