Claude Fable 5 addresses the issue of over-restricting biology questions, reducing false positives and cutting ‘fallbacks’ by approximately 85%.

Aug 10, 2026 20:50:00

Anthropic has updated the safety measures for its AI model ‘Claude Fable 5’ in the field of biology. They have reduced the number of cases where even low-risk questions are deemed to be subject to restriction, and the ‘fallback’ to switching to a different model for biology-related questions has decreased by approximately 85%.

Improving Fable 5 Safeguards\Anthropic

https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards

While Claude Fable 5 can handle advanced biological problems, the same capabilities could be misused for the development of biological weapons and other harmful applications. Because it is difficult to distinguish between safe and dangerous research, Anthropic broadly restricted almost all biology-related questions when Fable 5 was first released. Questions deemed to be restricted are then processed by ‘Claude Opus 5,’ which does not have the same high level of biological capabilities as Fable 5.

However, traditional safety measures were set quite cautiously, sometimes restricting even general questions about the test results or learning about biology. As a result of prioritizing not missing any potentially dangerous questions, there was an increase in ‘false positives’ where harmless questions were mistakenly detected.

This update significantly reduces the wide ‘safety margin’ that was in place when Fable 5 was first released, as shown in the diagram below. The range of safe questions that can be passed through has been widened, making it less likely that low-risk questions will be mistakenly restricted.

Anthropic rewrote the rules used for classification, updated the training data, and retrained the classifier that determines safety. Tests showed that biology-related fallbacks were reduced by approximately 85%, making Fable 5 easier to use in everyday health consultations, biology learning, and some clinical tasks performed by healthcare professionals.

However, even after the update, specialized questions deemed to be dual-use, such as those in virology, toxicology, and molecular design, will be switched to Opus 5. Anthropic states that it will continue to reduce false positives and develop mechanisms to ensure that researchers can safely utilize Fable 5’s advanced biological capabilities.