Claude Finds a CRISPR-Like Enzyme System in Phage DNA

Anthropic has revealed that it now runs its own molecular biology lab, and that Claude has already turned up something in it worth reporting. In a post published on 23 September, the company introduced a life sciences research group, formed in the spring of 2026, and described an enzyme system found in the DNA of bacteriophages — the viruses that infect bacteria — that no one had characterised before. Anthropic calls it ART, for array-associated reverse transcriptases, and says Claude found it "with only high-level direction from our scientists."

The search itself was run almost entirely by agents. Anthropic’s scientists wrote a prompt asking Claude to comb a massive database of DNA sequences for interesting new reverse transcriptases, enzymes that copy RNA into DNA. From there, roughly 950 Claude agents worked for 21 hours and used 210 million tokens. They gathered more than 200,000 reverse transcriptases, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling, each written up as a human-readable report. Anthropic says this kind of genome mining would take an expert weeks to months.

The find came when one agent was reading the raw sequence next to an odd-looking enzyme and noticed a regular pattern. "I can see by eye a tandem repeat array," it wrote in its notes, before asking itself whether it was "a CRISPR-like … repeat array?!" It then did what a scientist would: counted the repeats, measured their spacing, compared the layout with known systems, and searched the literature for any earlier report, before filing its case for human review. ART turned out to have three parts: the reverse transcriptase itself, an accessory protein of unknown function beside it, and a long array of evenly spaced repeat sequences.

That array is why the comparison with CRISPR is being made. A CRISPR array stores a bank of different RNA guides, which is what makes CRISPR-Cas systems programmable and turned them into gene-editing tools. Anthropic’s first lab experiments show the ART array is also expressed as a set of distinct short RNAs, which hints that something similar may be happening. The company is careful to say it does not yet know what ART does. What it can say is that this combination of features has only been seen together in a handful of other systems, all of them programmable and able to cut, copy or paste DNA.

It is also careful about what was new. The underlying enzyme, from a so-called jumbo phage, had been identified in earlier studies; Anthropic’s claim is that Claude appears to be the first to notice the array and the partner protein that define the system. Chief executive Dario Amodei wrote, as TechCrunch reported, that the discovery was made "mostly, though not entirely, by Claude," and acknowledged that a Stanford team had earlier described a system "in some ways similar" to this one. Feng Zhang, the MIT and Broad Institute scientist who helped pioneer CRISPR gene editing, reviewed the preprint and called the RNA-repeat arrays "genuinely intriguing" and worth further study.

The lab itself is deliberately modest. It sits in the Bay Area, works only at the lowest biosafety levels (BSL-1 and BSL-2), handles no pathogens that can infect humans, and every experiment is carried out by human scientists. Claude’s job is the part before the bench: reading the literature, reproducing known results to check its methods, proposing functions for unexplained genes, and then attacking its own proposals, which eliminates most of them. The team uses the same Claude Science and Claude Code tools any researcher can buy, plus an in-house harness that runs many sessions in parallel. Because Claude generates hypotheses by the thousand, the group now studies which ones its scientists judge worth testing and feeds that back into Claude’s instructions.

Anthropic has released a technical report and says more experiments are under way to work out ART’s function. Until they are, the discovery is a promising lead rather than a new tool. The more concrete result may be the method: a single prompt, one day of agent time and a short list of candidates that human biologists can afford to test. If that pattern holds up, the scarce resource in this kind of biology becomes bench time, not the search.