Claude's 21-Hour Search Turns Up a New CRISPR-Like Enzyme

Anthropic said Sept. 23 that an autonomous run of its Claude agents identified a previously uncharacterized enzyme system in bacteriophage DNA, built around a reverse transcriptase sitting beside a long array of evenly spaced DNA repeats resembling a CRISPR array. The company calls the system array-associated reverse transcriptases, or ART, and has not determined what it does.

The campaign ran for 21 hours across roughly 950 agent sessions, consuming 210 million tokens, according to Anthropic's post describing the work. Agents searched a database of about 1.9 billion protein clusters, narrowed the field to 3,564 candidate partner families, and produced 19 reports for human researchers to review before ART was identified as the standout finding.

What the repeats actually resemble

CRISPR systems use a bank of RNA guides, stored as an array of repeated DNA sequences, to make the associated enzyme programmable against a specific genetic target. Anthropic's post says the ART array follows the same layout: a reverse transcriptase enzyme, a partner gene beside it, and a repeat array that mirrors the structure CRISPR relies on. What the array actually guides the enzyme to do has not been established.

Anthropic released a preprint alongside the announcement, not yet peer-reviewed. Feng Zhang, a professor at MIT and the Broad Institute who helped pioneer CRISPR gene editing and was not part of the Anthropic team, reviewed the work and said the identification of RNA-repeat arrays tied to reverse transcriptases is "genuinely intriguing" and "merits further investigation."

What Anthropic has not shown

Anthropic's post does not claim ART is active or functional. The company has not demonstrated that the enzyme does anything to the RNA it appears to be associated with, and it says its own work to understand ART's primary function is ongoing. Anthropic also did not name which version of Claude ran the search, a detail its own post leaves out.

The scale of the search is itself the news, independent of whatever ART turns out to do. Running 950 agent sessions against a protein database that size, and reducing it to 19 human-reviewable reports, is the kind of workflow Anthropic has previously pointed to as evidence its models can contribute to research pipelines rather than just answer questions about them, a pattern also visible in how it has priced and pitched Opus 5.5 to enterprise customers. Coordinating hundreds of agent sessions without losing track of what earlier sessions already ruled out is the same problem projects like Hindsight's agent-memory framework are trying to solve for smaller, non-scientific workloads.

The numbers behind the search

1. Protein clusters searched1900 millions
2. Tokens consumed210 millions

Source: Anthropic, "Claude discovers a novel enzyme system with CRISPR-like repeats," accessed 2026-09-25

That ratio, 1.9 billion protein clusters searched against 210 million tokens spent, is itself a way of reading the funnel: the agents did not read each cluster in full. From the clusters, the search narrowed to 3,564 candidate partner families and then to 19 reports filed for human review. Of those, only one, ART, became the subject of Anthropic's announcement.

This is not the first time a lab has claimed an AI system found something in biology that a database search alone would not have surfaced, and it will not be the last. What distinguishes this case from a marketing claim is that Anthropic published the preprint, named an outside reviewer, and said plainly what remains unverified, rather than asserting ART works before anyone has tested it in a lab. Whether that restraint holds once ART's function is characterized, or whether Anthropic revises the claim, is the detail worth checking back on.