Anthropic Cuts Biology Filter False Positives 85%

Anthropic has cut the false-positive rate in Fable 5's biology safety filters by roughly 85 percent, a fix aimed at a problem scientists had been complaining about for months. Under the old system, nearly every biology-related question got flagged and bounced to Opus 5, a weaker model, regardless of whether the question had anything to do with actual danger. Researchers asking about lab results or symptom interpretation ran into the same wall as someone probing for weapons-grade information.

That wall is coming down, but only in part. Fable 5 can now handle a much wider range of biology queries directly, from reading lab panels to answering routine medical questions. Virology, toxicology, and molecular design are staying locked, and Anthropic isn't apologizing for it. The company argues these are the areas where a capable model could hand someone an edge they couldn't get anywhere else, and biological harm doesn't come with an off switch the way a cyberattack does.

Anthropic says it's building access programs to let vetted researchers eventually work with the restricted features. What's driving the caution, and whether it holds up against outside evidence, is where the real argument lives.