SAN FRANCISCO–A researcher trying to interpret a lab experiment on Claude, Anthropic’s AI assistant, would previously watch the model suddenly refuse the request or hand the task to a less capable model. Those interruptions, built into the system to prevent misuse of biological knowledge, frustrated users working on routine science. On Aug. 7, Anthropic said it had overhauled that machinery.
The company updated the biosafety controls on Claude Fable 5, its latest flagship model, by retraining the classifier that decides which biology-related requests require extra caution. In internal testing, the update cut the number of biology-related “fallback” events, in which the model downgrades its response or routes the task to a weaker model, by roughly 85 percent across the product’s platforms, the company said.
The change reflects a recalibration of how AI companies weigh two competing goals: keeping dangerous knowledge out of the wrong hands and letting legitimate users get full value from their tools. Anthropic said the revised system allows Fable 5 to handle more biology tasks directly, including everyday health questions, interpretation of experimental results, and biology education, without triggering a downgrade.
For medical professionals, the company said, the update expands what the model can support in clinical settings, where earlier safeguards sometimes blocked routine queries about drug interactions, lab values, or study designs. Clinicians testing the system reported fewer interruptions on standard questions, according to people familiar with the company’s evaluations.
The easing has limits. Anthropic said Fable 5 still switches to a lower-capability model for requests involving virology, toxicology, and molecular design, fields it classifies as carrying “dual-use” risk, where the same knowledge that supports legitimate research could enable the creation of harmful agents. Requests for drug-development work in those areas continue to route to the restricted tier.
The approach illustrates a broader industry argument about where the line should sit. Anthropic has positioned itself since its founding as the AI company most focused on safety, with a dedicated security team and a policy of releasing its most capable systems gradually. Competitors have pushed in the opposite direction at times, arguing that overly cautious filters erode the usefulness of their products for scientists and students.
The company framed the update as the product of better classification rather than relaxed standards. Its safety team trained the new classifier on a broader set of examples, including benign research workflows that earlier versions misread as risky, which reduced false positives without changing the policy governing genuinely dangerous requests, according to a person familiar with the work.
False positives have been a persistent complaint among professional users of AI tools in science. Researchers have documented cases in which models refused to answer questions about common laboratory techniques or disease mechanisms, forcing scientists to rephrase queries or switch tools mid-analysis. In biology education, instructors reported that students using the models hit unexplained walls on homework about genetics and microbiology.
The numbers Anthropic cited suggest the problem was widespread on Fable 5’s launch version. A roughly 85 percent reduction in fallback events, measured across platforms and user segments, implies that biology-related downgrades were happening frequently enough to shape the experience of a large slice of the scientific user base.
Industry observers said the update fits a pattern of AI labs tightening and loosening safeguards as they learn where the real risks sit. “The first generation of safety filters was built fast and conservative,” said one policy analyst who studies AI governance. “The second generation is built from data about what actually gets abused, and that tends to be a much narrower set of behaviors.”
Anthropic said it would continue to monitor how the relaxed filters behave in production, and that it retains the ability to restore stricter settings for specific models or regions if misuse appears. The company also said it is developing more granular controls that would let institutions set their own biosafety thresholds for Fable 5 deployments inside their organizations.
The commercial stakes are real. Anthropic competes with OpenAI and Google for scientific and medical customers, a segment that values reliability and is sensitive to tools that interrupt work. A model that refuses routine biology questions loses ground to competitors that handle them smoothly; a model that is too permissive risks reputational damage if its outputs are linked to a real-world incident.
The company has bet its roadmap on specialized capability tiers, and the Fable line is its flagship. Each version since the first Fable model has expanded the range of tasks the assistant can perform autonomously, and biosafety policy has been one of the few areas where Anthropic publicly documented restrictions.
For now, the update arrives quietly: a revised classifier, a lower false-positive rate, and a note to customers about what changed. The scientists who use the model daily will notice the difference in the absence of interruptions. The question that remains, and that regulators will keep asking, is whether the narrower restrictions still cover the requests that matter.
Anthropic said it briefed its safety advisory board on the change before rollout, and that external evaluators tested the revised filters against a set of hazardous-request benchmarks. The model passed the same suite of dangerous-capability tests it had cleared before, the company said, suggesting that the reduced fallback rate did not come at the cost of its guardrails. The company’s next documentation release will publish the details of the benchmark results.
