Anthropic released Claude Fable 5 this week, branding it the most capable AI model it has ever made publicly available, with particular praise for its biology knowledge. But there's a catch: the model flatly refuses to answer basic biology questions that a high school student could handle, routing them instead to the older Opus 4.8 model.

This isn't a capability gap. Fable knows the answers. Anthropic simply won't let it respond, by design. The company told The Verge the model's safeguards are "overly conservative" specifically to block bioweapons-related queries. The result is a system that declines prompts like "tell me about cell membranes," "what are mitochondria," or "how mRNA vaccines work." Even harmless medical questions—"what causes hay fever," "how does asthma medicine work"—get redirected.

Under the hood, this is a trade-off in the safety classifier pipeline. Anthropic has deployed Fable as a Mythos-class model, a family previously deemed too risky for public release due to cybersecurity capabilities. While the company warned extensively about cyber risks, biology is where the guardrails bite hardest. Fable's classifier appears to use broad keyword or embedding-based blocking that generates a high false-positive rate for any biology-adjacent query.

Architecturally, this is a multi-model routing pattern: Fable evaluates the prompt, and if the classifier flags it, the request falls through to Opus 4.8, which generally answers without issue. The routing adds latency and complexity, but Anthropic argues it's necessary. "We made this tradeoff so customers could benefit from the model's capabilities sooner without the risks," said spokesperson Paruul Maheshwary.

Interestingly, Fable is more permissive with chemistry and cybersecurity queries. It explained TNT basics (minus synthesis), chlorine gas use, and iPhone hardening. Only prompts like "how to make anthrax" triggered a full chat pause. The mitochondria refusal, however, is a textbook false positive—and one that highlights the operational cost of overly aggressive filtering.

Anthropic plans to eventually release unrestricted versions to the biology research community. For now, this restricted rollout sets a precedent for how safety engineering decisions directly shape user experience—and where the line between caution and obstruction gets drawn.

Source: The Verge