Claude autonomously found a new, CRISPR-like enzyme system -- once

Anthropic disclosed on September 23 that its newly formed life sciences research group -- established this spring to test whether general AI models can systematize biological discovery -- had Claude autonomously identify a previously uncharacterized enzyme system in bacteriophage DNA. Anthropic is calling it an array-associated reverse transcriptase (ART) system: a reverse transcriptase enzyme, a partner gene beside it, and a long array of evenly spaced DNA repeat sequences whose layout resembles a CRISPR array.

The scale of the search, and the honest caveat

Claude reached this result after autonomously searching a database of roughly 1.9 billion protein clusters over a 21.5-hour campaign spanning 949 agent sessions and 215.6 million tokens -- a genuinely large-scale autonomous research run. But Anthropic is explicit about two limits: it doesn't yet know what ART does, and when the same search was repeated ten additional times, none of the follow-up runs reproduced the find. That reproducibility gap is the load-bearing fact here, not a footnote -- a single successful run against a near-2-billion-entry search space is a promising signal, not a confirmed discovery, and Anthropic is presenting it as exactly that.

Why it matters beyond the specific finding

The result is Anthropic's first public evidence for the life sciences group's founding premise -- that a general-purpose model, not a narrow bio-specific one, can search unstructured biological data at a scale and speed no human team could match, and surface genuinely novel candidates. Whether ART turns out to be biologically significant is a separate, still-open question; what's demonstrated so far is the search capability, not yet a validated discovery.