Anthropic Cuts False Positives in Fable 5's Biology Filter by 85%, Keeps Virology and Toxicology Blocked

Anthropic has reduced false positives in Fable 5's biology safety classifier by approximately 85%, the company announced. The change addresses a problem that drew sharp criticism from scientists: nearly all biology-related queries were previously blocked outright and rerouted to Opus 5, a less capable model.

What changed

Under the old system, Fable 5's safety filter treated most biology questions as potentially risky, regardless of intent. A researcher asking about lab results or a patient trying to understand symptoms would get bounced to Opus 5 instead of receiving a direct answer from Fable 5. Anthropic says the retuned classifier now lets the large majority of benign biology queries through directly, including interpreting lab results, understanding symptoms, and answering general medical questions.

What hasn't changed

Restrictions remain in place for what Anthropic categorizes as "dual-use" topics: virology, toxicology, and molecular design. According to Anthropic, these are areas where Fable 5 could provide capabilities to bad actors that aren't readily available elsewhere. The company says it is building access programs to eventually let vetted researchers use these restricted features, though no timeline or eligibility criteria have been disclosed.

Anthropic's rationale

Anthropic argues that biological risks differ from cyber risks in a critical way: a released virus can't be shut down once it's out, and countermeasures take time to develop and deploy. The company points to several pieces of evidence it says support this concern, though most originate from outside sources rather than Anthropic's own research:

• An analysis by U.S. intelligence agencies on AI-enabled biological risks
• A review of ChatGPT conversations that reportedly included requests for bioweapon instructions
• Joint warnings issued by outside researchers about AI misuse in biology
• A Stanford project that used AI tools to help design fully synthetic viruses

Anthropic frames these as evidence that the threat is substantive rather than speculative caution. None of these claims are independently verified in this reporting, and the specifics of the intelligence analysis and the ChatGPT chat review have not been published in detail.

What this means

This update is a narrow but telling correction. Anthropic's original blanket restriction on biology queries was blunt enough to frustrate legitimate use cases, pushing users to a weaker model for routine medical and scientific questions. An 85% cut in false positives suggests the original classifier was badly miscalibrated, not just cautious.

The harder problem, dual-use research, remains unsolved. Virology, toxicology, and molecular design sit at the intersection of legitimate scientific inquiry and potential weapons development, and no classifier can perfectly separate the two based on query text alone. Anthropic's promised access program for researchers is the more consequential piece here: if it materializes with clear criteria, it could become a template other labs follow for handling dual-use science. If it stays vague, it risks looking like a permanent excuse for keeping capabilities locked rather than a genuine pathway to access. For now, the change mainly benefits everyday users with routine biology and health questions, while researchers working on genuinely dual-use topics are still waiting.