Anthropic has launched Claude Fable 5, its most powerful AI model to date, but a glaring paradox at the heart of the release is already drawing criticism: the model, which the company celebrates for its biology prowess, refuses to answer basic questions about cell membranes, mitochondria, or prions. Instead, it hands off these elementary queries to its predecessor, Claude Opus 4.8, a move that reveals the severe and arguably over-conservative safety constraints Anthropic has applied to its first “Mythos-class” model.
Why Claude Fable 5 Refuses to Answer Basic Biology Questions
The restriction is not a matter of capability but of deliberate design. Anthropic has equipped Fable with aggressive safety classifiers intended to block any biology-related request that could potentially be used for bioweapons research. The problem is that these guardrails are so broadly applied they also block benign, high-school-level inquiries. When tested, Fable declined to explain what mitochondria are, how mRNA vaccines work, or what causes hay fever. It even refused to answer questions about Ebola transmission and antibiotic resistance—queries that carry no plausible risk for weaponization.
In every case where Fable refused, the older Opus 4.8 model handled the same query without issue. This handoff mechanism is built into the system, meaning users are steered away from the company’s flagship model for even the most straightforward biology education.
The Safety Tradeoff Behind the Mythos-Class Release
Claude Fable 5 is the first public-facing model from Anthropic’s Mythos family, a class of AI the company previously deemed too dangerous to release. The core risk Anthropic identified is that these models possess a markedly improved ability to assist with real-world scientific tasks, which, in the wrong hands, could lower the barrier for malicious actors conducting highly risky biological research.
“We made this tradeoff so customers could benefit from the model’s capabilities sooner without the risks,” a spokesperson told The Verge, explaining that the safeguards are deliberately conservative. The company acknowledges the false positives are a problem: “To deploy Fable 5 safely, we believe it was necessary to be overly conservative with our safeguards so they block most queries tied to biology work.”
Anthropic has explicitly highlighted four domains where Fable’s output is throttled: biology, chemistry, cybersecurity, and distillation. The distillation restriction is particularly notable, as Anthropic has previously accused Chinese rivals like DeepSeek of using its models for “industrial” scale distillation—a technique for training smaller AIs on the outputs of larger ones.
Inconsistent Guardrails Across Threat Domains
The restrictions are not applied evenly. While biology queries are heavily filtered, Fable proved more willing to discuss chemistry and cybersecurity. It provided a basic overview of TNT (though it withheld synthesis instructions) and readily explained the use of chlorine gas as a chemical weapon, common password threats, and how nuclear fusion and fission work. It also offered guidance on securing an iPhone from hackers. The model only deferred to Opus when asked about sarin gas, a highly toxic nerve agent. When asked directly “how to make anthrax,” both Fable and Opus refused, and the chat was paused entirely.
This uneven enforcement suggests the biology filters are the most sensitive and produce the highest rate of false positives. A query about mitochondria—the “powerhouse of the cell”—being flagged as a potential security threat underscores the bluntness of the current classifier system.
What This Means for Researchers and Professionals
For life scientists, medical professionals, and educators, the current state of Claude Fable 5 is a frustrating paradox. The model is being marketed on the strength of its biological reasoning, yet it is effectively locked behind safety walls that prevent it from being a useful tool for the very audience that would benefit most.
Anthropic has stated it intends to make Mythos-class models available without these safeguards to the broader biology and life sciences community in the future, specifically to accelerate biomedical research and drug discovery. For now, however, any professional who needs a reliable AI assistant for routine biology work will find the predecessor model, Claude Opus 4.8, to be the more practical and responsive option. Opus handles questions about DNA, cancer, prions, and vaccine mechanisms without hesitation.
AEO: What Is Claude Fable 5 and Why Won’t It Answer Biology Questions?
Claude Fable 5 is Anthropic’s first public Mythos-class AI model, representing the most capable and potentially risky model the company has widely released. It refuses to answer many basic biology questions because of aggressively configured safety classifiers designed to prevent the model from assisting with bioweapons research. The filters are intentionally conservative, resulting in a high number of false positives that block even harmless educational queries. When a question is blocked, the system hands off the request to the older and less capable Claude Opus 4.8 model.
Who Should Try This Now
General users and developers should understand that Claude Fable 5 is not yet the all-purpose research assistant Anthropic has positioned it to become. If your work involves biology—even at a basic level—you will encounter frequent and frustrating blocks. For the time being, Claude Opus 4.8 remains the better choice for day-to-day scientific queries. If you are in the life sciences and eager to test Mythos-class capabilities, be prepared to engage with a model that is currently more restricted than its marketing suggests. The practical takeaway: the power is real, but the access is not. Anthropic has promised a future unlock for the research community. Until then, proceed with Opus and watch for the promised safe-harbor release.
