‘Access versus catastrophe’: Anthropic reopens biology on its top model

Anthropic said on Thursday it had updated the biology safeguards on Claude Fable 5, its most capable model. Everyday questions, interpreting lab results, understanding symptoms, learning biology, will now get through far more often. The company says the change cut biology-related “fallbacks” by about 85% across its products.

A fallback is what a user hits when a safeguard fires. Instead of answering, Fable 5 hands the request to Opus 5, an older and less capable model. At launch, Anthropic blocked almost all biology queries this way, accepting a flood of false positives. The over-blocking became a running joke.

Even now, one observer quipped, Fable 5 “still refuses to answer how babies are made”.

The friction has not been only theoretical. Through the tuning period, various users reported benign requests tripping the safeguards, with some sessions stopping mid-task. The hope now is that changes like this one settle things back to normal.

Why the guardrails were so tight

Anthropic’s reasoning is blunt. Fable 5 can now beat experts on some biology tasks, it says. The firm has built a team around chemical and biological weapons risk. Its own tests, it says, found the model could give a would-be bioweapon maker “significant uplift”, capabilities “they could not find anywhere else”.

Biology is hard to police because it is dual-use. Developing a live vaccine means growing the very pathogen you want to stop. The blood-pressure drug captopril came from toxins in snake venom. Anthropic says bad actors exploit that ambiguity to dress dangerous requests as ordinary research, and points to the US intelligence community’s 2026 threat assessment on state bioweapon programmes.

What actually changed

The fix is a rewrite, not a removal. Anthropic’s safeguards run on classifiers, small AI systems that flag risky requests. Over several weeks it rewrote the classifier’s “constitution”, the rules separating allowed from safeguarded content, with expert feedback, then retrained it. Crucially, those classifiers must also survive attempts to bypass them.

Plenty stays off-limits.

Fable 5 still falls back to Opus 5 for dual-use work like virology, toxicology and molecular design, so it is not yet usable for professional research or drug development. Anthropic says it will close that gap through trusted access for vetted researchers. A safety margin still blocks some low-risk questions out of caution.

A loosening, in a nervous week

The timing is striking.

Anthropic is widening biology access in the same stretch that scientists used AI to design working viruses and OpenAI’s agents broke out of a test sandbox. It is the paradox Anthropic keeps living: the more useful the model, the more Anthropic has to fence it. For now, the company is moving the fence, not taking it down.