이미지: Anthropic
Summary
- Anthropic retrained Fable 5's biology safety classifier, reducing unnecessary blocks (false positives) by about 85%.
- Everyday requests — general medical questions, understanding symptoms, educational biology content — are now handled directly by the more capable Fable 5.
- Specialized research domains with dual-use potential, such as virology, toxicology, and molecular design, remain blocked as before.
What changed
Anthropic announced on August 7 that it had substantially improved Claude Fable 5's biology safety classifier. The update cuts by roughly 85% the rate at which biology-related queries are "fallen back" to the less capable Opus 5 model. Anthropic, a company whose core mission is AI safety research, aims to build AI systems that are trustworthy, interpretable, and controllable.
The practical beneficiaries of this change are users with everyday medical and educational needs. Requests such as interpreting test results, understanding symptoms, or learning biology in an educational context are now handled directly by Fable 5. Medical professionals will also be able to draw on more support from Fable 5 in their clinical work.
Why the block was so broad from the start
Anthropic explained that when Fable 5 launched, it deliberately blocked nearly all biology-related queries. This was because the model had reached a level in some highly complex biology tasks that surpassed human experts. According to the company's capability evaluations, Fable 5 could provide malicious actors with a "meaningful uplift" — capabilities they could not otherwise obtain.
A key concern is that, unlike cyberattacks, biological risks are difficult to reverse once they manifest. A released virus cannot be shut down remotely, and developing countermeasures takes time. Anthropic cited the U.S. intelligence community's 2026 Annual Threat Assessment, noting that advances in biotechnology — including synthetic biology and genome editing — could give rise to new biological threats, and that some state actors likely maintain active offensive biological and chemical weapons programs.
Another reason for maintaining such a broad block was the difficulty of distinguishing beneficial biology research from harmful research. For instance, developing live vaccines requires scientists to culture the very pathogens they are trying to prevent, and the hypertension drug captopril was developed by isolating a compound from snake venom that sharply lowers blood pressure. Therapeutic research and the production of hazardous substances often occur within the same process.
How the classifier works
Anthropic's primary tool for managing biological risk is a safety classifier — a smaller, automated AI system that detects when Fable 5 is attempting to perform certain biology-related tasks or generate harmful output. When the classifier is triggered, the user's request is routed to Opus 5, which lacks the same level of biology capability.
To achieve this improvement, Anthropic spent weeks completely rewriting the "classifier constitution" — the set of rules distinguishing permitted from blocked content. It was redesigned to more finely categorize benign use cases, incorporating feedback from both internal and external expert groups. Anthropic then built new training data based on the revised constitution and retrained the classifier, verifying that it still functioned against harmful or dual-use content while allowing a broader range of benign uses.
The impact of the update varies by product surface. Overall fallback is expected to drop by about 67% on Claude.ai, about 55% on Cowork, about 17% on Claude Code, and about 7% on the Claude Platform.
What's still blocked
Specialized research domains with dual-use potential remain blocked even after this update. Queries related to specialized biology research and drug development — including virology, toxicology, and molecular design — will continue to be routed to Opus 5. Anthropic explicitly stated that, as a result, Fable 5 still cannot be used for specialized biology research or drug development purposes.
Additionally, false positives will not disappear entirely for requests that fall within the classifier's safety margin — cases where risk is very low but the classifier is still triggered. Anthropic acknowledged this limitation and said it is building a program that would let researchers access limited functionality through "trusted access pathways." The plan is to give specialized researchers access to Fable 5's full capabilities without dual-use restrictions, though specific timelines or conditions have not yet been disclosed.
