Claude Fable won’t answer basic biology questions
Anthropic’s new Claude Fable 5 blocks many basic biology prompts, even harmless ones, and falls back to Opus 4.8 instead.
Intelligence analysis by GPT-5.4 Mini
Anthropic says Claude Fable 5 is powerful enough to create real safety risk, so it uses conservative biology filters that block many ordinary questions. The result is a model that can answer some science prompts, but often refuses basic biology and hands the query to an older model.
Anthropic gave its new AI a very strict rulebook for biology, like a library that locks away even some normal science books so nobody can grab the dangerous ones. That means it sometimes refuses easy questions too, even when a student would be allowed to ask them.
Analysis
What happened
The Verge reports that Anthropic’s newly released Claude Fable 5, described by the company as its most powerful widely available model, refuses a surprising number of basic biology questions. Prompts like “what are mitochondria,” “tell me about cell membranes,” and “how mRNA vaccines work” were blocked, even though they are ordinary educational questions.
When Fable refused, Anthropic’s older Claude Opus 4.8 usually answered instead. That makes the restriction especially visible: the model does not appear unable to handle the topic, but is being held back by policy.
Why Anthropic did it
Anthropic told The Verge the restraint is deliberate. The company says Fable’s new capabilities raise the risk that malicious actors could use it for highly risky biological research. In response, Anthropic says it chose to be “overly conservative” with safeguards so they would block most biology-related queries.
The company has previously said it would throttle Fable across four areas: chemistry, biology, cybersecurity, and distillation. In this test, chemistry and cybersecurity looked less restricted than biology. The model could give broad explanations about TNT, chlorine gas as a weapon, password threats, nuclear fusion and fission, and iPhone security, but it still refused some dangerous prompts such as sarin gas and anthrax-related requests.
The tradeoff
Anthropic frames the release as a compromise: ship the model now, but limit risky biology behavior until detection improves. The company also says it wants future Mythos-class models to be available without these safeguards to the broader biology and life sciences community, so they can help with biomedical research and drug discovery.
The downside is obvious in the current release. The Verge found at least one false positive, where a mitochondria question was blocked despite being harmless. That suggests the safety filters may be protecting against misuse while also blocking legitimate learning and research.
Key points
- Claude Fable 5 blocks many basic biology questions and often hands them to Opus 4.8.
- Anthropic says the biology filters are intentionally conservative because of bioweapon risk.
- Some dangerous prompts still remain blocked, but the filters also appear to create false positives.
- Anthropic says it wants future Mythos-class models to support biomedical research and drug discovery.
- The article highlights the tradeoff between safety controls and usefulness for legitimate work.
If Anthropic tunes the filters better, it could keep the model safe while letting students and researchers ask normal biology questions. The company also says future versions may be opened up more for biomedical research and drug discovery, which could make the model more useful to scientists.
The current setup may block too many harmless questions, making the model frustrating to use for education or research. If the safeguards stay this conservative, users may keep being pushed to older models, and the biology community may not get the capabilities Anthropic says it wants to share.
