Anthropic loosens Claude Fable 5's biology filter, cutting blocked questions by 85 percent

If you have ever asked Claude Fable 5 to help you read a blood test result and got a noticeably weaker answer than you expected, this is why. Anthropic said on Thursday that it has retrained the safety system that screens biology questions, and the change cuts biology-related “fallbacks” by roughly 85 percent across its products.

A fallback is what happens when Fable 5, the company’s most capable model, decides a request touches a sensitive area and quietly hands it to Opus 5 instead. Opus 5 is a strong model, but it does not have Fable’s biology capability, so the answer you get is different from the one you asked for. Anthropic says the total number of fallbacks of every kind should now drop by about 67 percent on Claude.ai, 55 percent in Cowork, 17 percent in Claude Code and 7 percent on the Claude Platform.

The gatekeeper here is a safety classifier: a smaller, faster AI system whose only job is to look at a request and decide whether it falls inside a protected category. Anthropic rewrote the rulebook that classifier follows, had internal and external experts review it, generated fresh training examples and retrained the thing. The result is a boundary that sits in a more sensible place. Everyday health questions, symptom explanations and school biology now get through. Virology, toxicology and molecular design still do not.

Why the filter was so aggressive in the first place

Anthropic is candid that the original setting was deliberately blunt. Its own capability assessments found Fable 5 could give a malicious actor meaningful help toward a biological weapon, which is the kind of risk you cannot walk back after the fact. So the company shipped the model with almost all biology blocked, accepted that legitimate users would be caught in the net, and kept working on the classifier afterward. The alternative was holding the whole model back for weeks or months.

The hard part is that biology genuinely resists clean lines. Anthropic gives a nice example: developing a live vaccine means growing the very pathogen you are trying to stop. The blood pressure drug captopril came out of isolating the toxic part of snake venom. Beneficial and dangerous work often look identical from the outside, and people acting in bad faith know how to dress one up as the other.

A fair caveat: false positives have not disappeared. Anthropic says some low-risk requests will still trip the classifier, on purpose, as a safety margin. And Fable remains unusable for professional biology research or drug development until the company builds what it calls trusted access pathways.

What this means for you: if you use Fable 5 and have been bouncing off biology questions, try again. Interpreting lab results, understanding symptoms and general learning should now work much more smoothly. If you are a professional in life sciences, the useful signal is different: the door is still shut for you, but Anthropic has now said publicly that opening it is the plan. And for everyone else, this is a clear look at a tradeoff every AI company is making quietly, between blocking too much and blocking too little.

Sources

• Improving Fable 5’s biology safeguards, Anthropic
• Anthropic Newsroom