Anthropic has updated the biology safeguards built into Claude Fable 5[1]. The company says fallbacks, where a biology-related question is quietly rerouted to a less capable model, dropped by roughly 85 percent in its own testing. Everyday health and study questions that used to get caught in the filter should now go through.

Why Almost Every Biology Question Was Blocked

Fable 5 shipped with nearly all biology queries blocked, and Anthropic describes that as a deliberate choice rather than an oversight[1].

The reason is the model itself. By Anthropic's own assessments, Fable 5 can outperform experts on some highly complex biological tasks and provide operational support on others[1]. That makes it a powerful partner for a researcher developing a new treatment, but the same capability could help a malicious actor build a biological weapon. Anthropic states plainly that Fable 5 could give such an actor capabilities they could not find anywhere else[1].

The hard part is that beneficial and harmful uses are genuinely difficult to tell apart. Researching a cure sometimes requires handling the very compounds that cause the disease. Anthropic points to live vaccines, which require growing the same pathogen a scientist is trying to prevent[1]. The hypertension drug captopril also began with isolating components of snake venom that crash blood pressure in humans[1].

That ambiguity is convenient for anyone trying to hide their intent, because dangerous work can be dressed up as ordinary research. Anthropic cites the US Intelligence Community's 2026 Annual Threat Assessment, which notes that advances in synthetic biology and genomic editing could lead to novel biological threats, and that several state actors likely maintain active offensive biological and chemical weapons programs[1].

So the company shipped with the biology door held shut, getting the model to users in other domains first. A high false positive rate was priced in, on the view that it beat delaying general access by weeks or months[1].

Rewriting the Classifier's Constitution

Fable 5's biology guardrails run on safety classifiers, smaller automated systems that detect when the model is asked to perform a safeguarded biology task or produce a harmful output[1]. When one fires, the request is rerouted to Opus 5, which lacks Fable 5's level of biological capability and therefore offers less assistance to a bad actor. That reroute is the fallback users see.

Building a precise classifier is not simple. It has to learn the boundary between in-scope and out-of-scope content, hold down both false positives and false negatives, and resist jailbreak attempts[1]. Anthropic runs similar classifiers in cybersecurity and has previously described how they work[2].

What changed this time is the classifier's constitution, the rule set that separates safeguarded content from allowed content[1]. Anthropic spent several weeks rewriting it, carving out benign uses in detail, and solicited feedback from a range of internal and external experts. It then generated new training data from the revised constitution, retrained the classifier, and verified that it still triggers on harmful and dual-use research content while permitting a far wider set of benign requests[1].

What Loosened, and Where the Line Still Sits

In practice, the change shows up in ordinary health and education questions. Interpreting lab results, understanding symptoms, and studying biology in an educational context should now hit far fewer fallbacks[1]. Healthcare professionals will also get more support from Fable 5 on clinical tasks.

Total fallbacks, including those unrelated to biology, are expected to fall as well. Anthropic estimates a drop of roughly 67 percent on Claude.ai, 55 percent on Cowork, 17 percent on Claude Code, and 7 percent on the Claude Platform[1]. The spread between products tracks how much biology traffic each one sees.

The categories that did not loosen are stated just as clearly. Requests judged to be dual-use, including virology, toxicology, and molecular design, still fall back to Opus 5[1]. In other words, Fable 5 is not yet usable as-is for professional biology research or drug development. Anthropic says it intends to close that gap through trusted access pathways that give vetted researchers frontier capability[1].

The company is also explicit that false positives will not disappear entirely[1]. Very low-risk requests that land inside the classifier's safety margin can still trigger it.

Summary

Anthropic reworked Claude Fable 5's biology safeguards by rewriting the classifier constitution, cutting biology-related fallbacks by about 85 percent. Everyday uses such as interpreting lab results or studying biology should see far fewer reroutes to a weaker model. Virology, toxicology, and molecular design remain out of scope, and full research access waits on trusted access pathways. It is a concrete, numbered example of opening a deliberately narrow door step by step, using real data and expert review.

Source [1] https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards

Source [2] https://www.anthropic.com/news/fable-safeguards-jailbreak-framework