Anthropic reduces AI false positives in Fable 5

Anthropic has updated Claude Fable 5's biological safeguards, reducing unnecessary blocks on biology queries by approximately 85%. The change allows the model to answer more legitimate questions instead of automatically sending them to a less capable system.

That diversion, called a fallback, occurs when safety filters detect a potentially risky query and pass the request to Opus 5. The alternative model still responds, but has fewer biological capabilities than Fable 5.

More answers for health and education

With the update, you should encounter fewer interruptions when asking everyday questions about health or learning. Anthropic mentions examples such as:

• Interpreting lab results.
• Understanding symptoms.
• Learning biology concepts.
• Receiving support with certain clinical tasks.

The improvement also applies to healthcare professionals, although it does not mean Fable 5 is unrestricted for every medical or scientific use.

In Anthropic's tests, biological requests that ended in a fallback fell by about 85% across the different product surfaces. The company also expects the total number of diversions for any reason to decrease: by 67% in Claude.ai, 55% in Cowork, 17% in Claude Code and 7% in Claude Platform.

Why Fable 5 blocked so much biology

Fable 5 can outperform experts on some complex biological tasks and provide operational support on others. That capability can help research treatments, but it could also help someone develop biological weapons.

The problem is that many useful and dangerous activities look similar. Researching a vaccine, for example, may require working with the same pathogen it is meant to prevent. Developing captopril, a medication for high blood pressure, also involved studying toxic components of snake venom.

That is why Anthropic considers biology a dual-use area: a technology can have beneficial or harmful applications, and it is not always easy to distinguish between them based on a question.

How the new safeguards work

Fable 5 uses safety classifiers, smaller AI systems that analyze whether a request falls into a restricted biological category or whether the response could be dangerous. When they detect risk, the request is redirected to Opus 5.

The challenge is reducing two opposing errors:

• False positives: blocking a legitimate, low-risk question.
• False negatives: allowing a request through when it should be restricted.

Anthropic says it rewrote the classifier's internal rules, added more detailed descriptions of permitted uses and retrained it with new data. It also consulted internal and external experts to verify that the filter continued to detect harmful and dual-use content.

The practical result is simple: ordinary biology questions are more likely to receive a response directly from Fable 5 instead of being treated as potentially dangerous simply because they mention a sensitive topic.

Professional research still has limits

The update does not open Fable 5 to advanced biological research. Anthropic continues to block professional dual-use queries related to virology, toxicology, molecular design and drug development.

The company says it is working on controlled access routes for authorized researchers. The idea is to offer the most advanced capabilities to professionals who can demonstrate a legitimate use, without exposing those same tools to malicious actors.

For you, the change means fewer blocks when using Fable 5 to learn, review health information or work on permitted clinical tasks. What remains to be seen is whether the new filters can preserve that convenience without allowing dangerous requests through. In biology, improving AI is not only about making it more capable, but also about deciding precisely when it should respond and when it should stop.