Claude Fable 5 blocks basic biology questions by design

Anthropic says its newest Mythos-class model, Claude Fable 5, is its most powerful widely available AI yet—but when users ask for straightforward biology explanations like cell membranes or mitochondria, it often hands the request to Claude Opus 4.8 or refuses

Anthropic calls Claude Fable 5 its most powerful AI model it has ever made widely available. But in practice, it can feel like walking up to a tutor who refuses to answer the simplest chapter.

In testing, Claude Fable 5 repeatedly declined basic biology questions that a high schooler would handle with ease. It would not “tell me about cell membranes.” It refused “what are mitochondria.” It also declined to explain “what is a prion. ” the proteinaceous particles behind mad cow disease. and it wouldn’t answer “how mRNA vaccines work.”.

When Fable wouldn’t respond, it often didn’t just shut the door—it passed the question along to Anthropic’s former flagship model, Claude Opus 4.8, which generally answered the prompts well.

The reason, Anthropic says, isn’t that Fable lacks the knowledge. It’s that the biology restrictions are part of the model’s design.

Fable 5 is described as a public-facing, Mythos-class model. Anthropic has previously said models in the Mythos family are capable at cybersecurity tasks. and it had warned that releasing them publicly could be too dangerous. In this release. the safety tradeoffs appear most visible in biology: the guardrails are stricter there than in other areas. even when the requests look objectively harmless.

A range of medical or everyday biology questions ran into the same wall. Fable would not answer “what causes hay fever,” explain how asthma medicine works, explain how antibiotic resistance arises, or tell users what Ebola is and how it spreads.

At the same time, some broad topics did get through. Fable sometimes answered questions like “what is cancer” and “what is DNA.”

Anthropic is explicit about the philosophy behind the filters. In a statement to The Verge, spokesperson Paruul Maheshwary said the company made “this tradeoff so customers could benefit from the model’s capabilities sooner without the risks.”

Maheshwary said Anthropic uses classifiers to block model help for bioweapons-related requests. To deploy Claude Fable 5 safely, she said the company believes it was “necessary to be overly conservative with our safeguards so they block most queries tied to biology work.”

She also tied the approach directly to the launch of Claude Fable 5 itself. describing it as Anthropic’s “first Mythos-class model” and saying the company believes models now have greater ability to accomplish real-world scientific tasks—along with the possibility for malicious actors to use the technology for “highly risky biological research.”.

Anthropic has previously highlighted four key areas where it would throttle Fable’s responses for safety: chemistry. biology. cybersecurity. and distillation. a technique for training smaller AIs using the outputs of larger ones. The company has also accused Chinese rivals like DeepSeek of using distillation on an “industrial” scale.

In the same hands-on testing, the shape of the restrictions seemed uneven. Chemistry and cybersecurity questions were more likely to be answered than basic biology. For example. Fable gave a basic overview of the explosive TNT. while withholding synthesis instructions “for obvious reasons.” It also readily answered questions about the use of chlorine gas as a chemical weapon. common password threats. nuclear fusion and fission. and how to secure an iPhone from hackers.

Even within that comparatively open lane, limits remained. Fable deferred to Opus 4.8 when asked about sarin gas, and both Fable and Opus refused the prompt “how to make anthrax.” At one point, Claude paused the chat entirely.

In the biology case. the refusal patterns felt—at least from a user’s perspective—like the risk controls may have overreached. The refusal to explain mitochondria, or what a prion is, may not look like an obvious pathway to bioweapons work. Anthropic acknowledges the possibility of false positives and says it is working to improve detection.

Maheshwary said Anthropic is “working hard to improve its detection and reduce the false positives.” She also said the company intends to make Mythos-class models available without these safeguards to the broader biology and life sciences community. so the capabilities can be used to accelerate biomedical research and drug discovery.

What’s less clear is whether this restricted-release approach will become the default for future models. Anthropic did not answer questions about whether this kind of throttled availability will become the new norm.

For now, Claude Fable 5 arrives with a paradox baked in: it’s positioned as a major scientific leap forward, yet some of the most basic lessons in biology are the ones it won’t reliably teach.

Claude Fable 5 Claude Opus 4.8 Anthropic AI model safety biology safeguards bioweapons mitochondria mRNA vaccines prions cybersecurity distillation