Anthropic announced a new life sciences research group on 23 Sep 2026, with a lab in the Bay Area and early results to show for it. Claude agents mined a database of DNA sequences for interesting reverse transcriptases (RTs, enzymes that copy RNA into DNA) and surfaced one system no one had described: an RT sitting next to an array of evenly spaced DNA repeats, with a partner gene beside it. They call it array-associated reverse transcriptase, or ART.
The campaign scale is the part worth keeping straight. Anthropic says they gave Claude a prompt to search for new RT examples, and the agents gathered over 200,000 RTs, picked out thousands of candidate systems, and analyzed the most compelling into human-readable reports. Roughly 950 agents ran for about 21 hours and 210 million tokens. Their involvement was the initial prompt and the lab work; the agents did the combing and the judging.
The discovery itself happened in the transcripts, which the pre-print includes. It came from a side branch of the campaign: while the census scored partner-gene associations, workers also flagged unusual features of individual RTs and their flanking DNA for follow-up. One agent, reading raw DNA next to an odd-looking RT, wrote: "[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!" It then did what the announcement says a human scientist would do: counted the repeats, measured their spacing, compared the layout against known RT systems, searched the literature for any prior report of the pattern, and filed a report for human review.
The underlying RT had been identified in earlier studies, in a jumbo phage (the large viruses that infect bacteria). What Claude appears to be the first to notice is the system's defining features: the associated repeat array and the accessory protein of unknown function. The repeat layout resembles a CRISPR array, and their first experiments show the ART array is expressed as a set of distinct short RNAs, suggesting something analogous may be at play. The function is still unknown.
The pre-print goes further than the announcement, with the numbers a bit more precise: 119 tasks, 949 sessions, 77 agent-hours, 215.6 million tokens. It includes fixed-input benchmarks across seven Claude models, and a finding the announcement does not mention: more information and tooling hurt the models' ability to describe the repeat array, while simply reading the DNA into context helped. Recognition rose from 29% to 76-96% of attempts as the models read longer contiguous stretches of the sequence.
Feng Zhang, one of the pioneers of CRISPR genome editing, reviewed the pre-print and called the RNA-repeat arrays associated with reverse transcriptases "genuinely intriguing" and worth further investigation.
The lab itself is conventional: BSL-1 and BSL-2 work, no human pathogens, all lab work performed by human scientists. The team works in Claude Science and Claude Code, the same tools any scientist can use, plus a harness of their own for parallel sessions. One workflow detail they flag: because Claude generates hypotheses prolifically, the hypotheses have become an object of study. Hundreds to thousands of candidate reports per campaign, with most eliminated at the review stage, and what distinguishes the ones worth testing goes back into the instructions.
So the verified claim is narrower than "Claude discovered an enzyme system." The RT existed and was known. What is new is the association: the repeat array, the partner gene, the short RNA expression, and the observation that models reading raw sequence can notice what a pipeline of association filters would not surface. The pre-print's own discussion is the honest read: no enzymatic activity demonstrated, no substrate shown, function unknown.