Anthropic has launched Claude Fable 5, which it touts as the most powerful AI model available to the public to date, showcasing impressive capabilities particularly in the realm of biology. However, users may experience some frustration, as the model refrains from answering a number of basic biology questions—seemingly the type a high school student could handle—redirecting them instead to the older Claude Opus 4.8.
This restriction is intentional; Anthropic has designed Fable 5 with specific safeguards to limit its responses, particularly in the biological domain, to prevent possible misuse. Although the model itself possesses the knowledge necessary to answer these queries, Anthropic deliberately constrains its capabilities as a precautionary measure.
Fable is categorized as a Mythos-class model, a designation that reflects its advanced skills, especially in cybersecurity tasks. Notably, the company has previously raised concerns about the potential for misuse of such powerful AI technologies. However, while much of the safety discourse has revolved around cybersecurity, the limitations concerning biological inquiries stand out as particularly stringent.
In tests conducted by users, Fable consistently declined to address straightforward biology questions such as “Tell me about cell membranes,” “What are mitochondria?” and “What is a prion?”—the latter being associated with mad cow disease. Even more universally accepted questions like “What causes hay fever?” and “How does asthma medication work?” went unanswered. On rare occasions, the model did provide responses to questions surrounding “What is cancer?” and “What is DNA?”, but instances of refusal were far more common.
Anthropic explains that these stringent filters have been put in place for safety reasons. According to spokesperson Paruul Maheshwary, the focus has been on mitigating risks associated with potential bioweapon inquiries. “To deploy Fable 5 safely, we believe it was necessary to be overly conservative with our safeguards,” Maheshwary stated. The main apprehension concerns the potential for malicious actors to exploit such powerful models for hazardous biological research.
In addition to biology, Anthropic identified four primary domains where it limits Fable’s capabilities: chemistry, biology, cybersecurity, and distillation—the latter referring to a method of training smaller AIs using outputs from larger ones. The company has voiced concerns that competitors might be leveraging AI models akin to its own for industrial-scale distillation.
While Fable effectively handles questions in chemistry and cybersecurity, with certain limitations, encounters with refusal were more pronounced regarding biological inquiries. For instance, it could discuss general information about the chemical TNT but withheld synthesis instructions, and it answered questions related to chemical weapons and cybersecurity essentials. However, when strictly dangerous subjects arose, such as inquiries into sarin gas or anthrax, both Fable and Opus 4.8 declined to respond altogether.
Despite the current constraints, Anthropic’s goal is to enhance its detection systems and reduce false positives in response refusals. The company also expressed intent to eventually make Mythos-class models available without such stringent restrictions, allowing researchers in biology and life sciences to harness these technologies for advancements in biomedical studies and drug development.
As of now, questions remain about whether this restricted approach will characterize future model releases from Anthropic.
