Date: August 7, 2026
Anthropic announced updates to the safety classifiers that govern Claude Fable 5’s handling of biology-related queries, aiming to substantially reduce false-positive “fallbacks” that reroute requests to a less capable model. In their testing the change reduced biology-related fallbacks by about 85% across product surfaces.
What users will notice
Users should see far fewer fallbacks on everyday health and educational questions — for example, interpreting lab results, understanding symptoms, and learning biology in an educational context. Healthcare professionals can expect more assistance from Fable 5 on clinical tasks. However, Anthropic says Fable 5 will continue to fall back to Opus 5 for requests the company deems dual-use (including virology, toxicology, and molecular design), so it is not yet usable for professional biology research and drug development.
Why strong biology safeguards were built
Anthropic’s goal is to get Fable 5’s frontier capabilities to as many users as possible, but this requires managing increased risks. Their capability assessments indicate that Fable 5 can outperform experts on some complex biological tasks and provide operational support on others — capabilities that could benefit researchers developing new medical treatments but could also be misused by malicious actors, for example to develop biological weapons.
The company notes the difficulty of distinguishing beneficial from harmful biology use: some legitimate research requires producing dangerous compounds (for instance, live vaccines or certain medicines). Anthropic also cites the US Intelligence Community’s 2026 Annual Threat Assessment, which warns that advances in biotechnology — including synthetic biology and genomic editing — could lead to novel biological threats and that several state actors likely maintain offensive biological and chemical programs.
Because of these dual-use concerns, Anthropic initially launched Fable 5 with almost all biology queries blocked, accepting a high rate of false positives so the model could be available for other domains while more precise safeguards were developed.
How the biology safeguards work
A core protection is the safety classifier: smaller automated AI systems that detect when Fable 5 is asked to perform a safeguarded biology task or might produce a harmful output. When a classifier fires, the request is routed to Opus 5, a capable but less biologically powerful model, producing the visible fallback.
Designing precise, robust classifiers is challenging. They must rapidly and consistently distinguish between content that is “in scope” (safeguarded) and “out of scope” (allowed), minimizing false positives and false negatives while resisting attempts to bypass them (so-called jailbreaks).
What Anthropic changed
At launch, Anthropic used a very broad classifier that triggered on many requests, including likely benign ones, enabling general access to Fable 5 but causing many blocked queries. Over the past weeks they rewrote the classifier’s constitution — a rule set to help the classifier discern safeguarded from allowed content — and sought feedback from diverse internal and external experts. They then produced updated training data based on that constitution, retrained the classifier, and verified that the new classifier still triggers on harmful and dual-use research content while permitting a wider range of benign uses.
As a result, the classifier now fires for far fewer benign biology-related requests. Anthropic describes this as a rightward shift of the classifier boundary: more benign requests fall on the allowed side while harmful and dual-use content continues to be blocked.
Conclusions and next steps
Anthropic acknowledges more work remains to refine safeguards. False positives will inevitably persist — some low-risk requests will remain inside the classifier’s safety margin and be blocked. The company reiterates that dual-use professional biology and drug development queries will stay blocked due to potential risks, and it remains committed to creating safe, scalable “trusted access” pathways so researchers can use its most capable models.
Anthropic asks users to continue providing feedback to further improve the classifiers.
Numbers
- The update reduced biology-related fallbacks by about 85% across product surfaces in Anthropic’s testing.
- A footnote states expected reductions in total fallbacks by product: approximately 67% on Claude.ai, 55% on Cowork, 17% on Claude Code, and 7% on the Claude Platform.
