FABLE 5 WONT TELL YOU WHAT A MITOCHONDRION IS

Anthropic released Claude Fable 5 this week as its most powerful AI model available to the public. It is also the most restricted.

Fable, the first public-facing model from Anthropic's Mythos-class lineup, refuses to answer basic biology questions. Not because it does not know the answers. By design, it will not give them. When asked about cell membranes, mitochondria, prions, or mRNA vaccines, Fable declines and redirects the query to Claude Opus 4.8, the company's previous flagship. The refusal applies to medical queries too: hay fever, asthma medication, antibiotic resistance, and Ebola all get blocked.

"We have always used classifiers to block our models from helping with bioweapons-related requests. 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."

The company's head of safety, Paruul Maheshwary, framed the restrictions in stark terms. "With the launch of Claude Fable 5, our first Mythos-class model, we believe models now have a greater ability to accomplish real-world scientific tasks and for malicious actors to potentially use our models for highly risky biological research"

"GUARDRAILS" APPARENTLY

Anthropic identified four areas where Fable would throttle its responses: chemistry, biology, cybersecurity, and distillation. The biology filters are the most aggressive and the most noticeable. Fable answered basic chemistry questions readily, including a general overview of TNT while withholding synthesis steps. It discussed chlorine gas as a chemical weapon, password attack vectors, nuclear fusion and fission, and iPhone security without hesitation. When asked about sarin gas, Fable deferred to Opus. Both models refused queries about manufacturing anthrax.

The asymmetry is deliberate. Chemistry and cybersecurity questions get substantially more latitude than biology ones. The rationale, according to Maheshwary, is bioweapons. Anthropic considers biological research the primary risk vector for its most capable models, and it chose to be conservatively broad in filtering, knowing that some legitimate queries would get caught in the net.

RIDDLED WITH FALSE POSITIVES

The mitochondrial refusal is a case in point. Anthropic acknowledges it as a false positive: mitochondria are fundamental to cell biology and should not trigger a guard. The company is working to sharpen its detection and reduce these errors, but the underlying tension is structural. Overblocking is the chosen trade-off.

Some queries do get through. Questions about cancer and DNA occasionally pass the filter. This suggests the guardrails are not uniformly applied across all biology-adjacent topics, which raises the question of where the boundary actually sits and whether it is drawn in the right place.

ANTHROPIC HAS ALREADY SIGNALED WHAT COMES NEXT

"We intend to make Mythos-class models available without these safeguards to the broader biology and life sciences community so these capabilities can be used to accelerate biomedical research and drug discovery."

The plan creates a two-tier system: a locked-down model for the general public and an unlocked version for credentialed researchers in biology and life sciences. The company did not say whether this restricted-release approach will become standard for future Mythos-class models.

The arrangement mirrors a broader industry dynamic. Anthropic has accused Chinese AI company DeepSeek of using distillation on its models at industrial scale, essentially extracting capabilities from Anthropic's systems into competing products. The chemistry and cybersecurity throttles may be partly calibrated against this threat: limiting what a distilled version of Fable could reproduce.

WHAT THIS MEANS FOR USERS

For the foreseeable future, anyone asking Fable basic biology questions will hit a wall. The model knows the material, but it will not share it. Users get redirected to an older model that answers more freely, which creates an odd experience: the newest, most capable model is less useful for a wide range of legitimate scientific questions than its predecessor.

The pattern raises a structural question about how AI companies handle capability and safety trade-offs at the frontier. Anthropic built its most powerful model and then deliberately limited it for a broad category of queries. It is a visible admission that the model is capable enough to be dangerous, and that the company does not yet trust deployment at full strength.

Whether offering restricted access to researchers is a sustainable model, or whether it simply defers the same risk to a smaller audience, is a question that will define the next phase of frontier AI governance.

P.S. - It's powerhouse of the cell.