Anthropic says Claude found a never-before-described enzyme system: what's in the preprint and what's missing

According to Anthropic's description, the system has three components: a reverse transcriptase, a partner gene and a series of regularly spaced DNA repeats. It is this last arrangement that the company says resembles a CRISPR array — and that sentence deserves a pause, because it is the one most open to misreading. The stated similarity concerns the structure of the repeats, not their function: no ability to cut or edit DNA has been demonstrated in the published material. In fact, Anthropic writes that the function of the ART system and its associated accessory protein is not yet known. The first experimental result reported is that the array is expressed as a set of distinct short RNAs; according to The Next Web, citing the technical report, these RNAs made up about 8% of a Staphylococcus phage's RNA fifteen minutes after infection. The reverse transcriptase itself, found in a so-called jumbo phage, was already known from earlier studies: what the company claims as new is spotting the features that define the system around it.

The numbers behind the agentic work are the part that will travel furthest, and it is worth keeping the two versions apart. Anthropic's blog speaks of roughly 950 Claude agents, 21 hours of research, 210 million tokens, more than 200,000 reverse transcriptases collected, about 3,500 candidate systems and the 20 most promising taken forward. The precise figures — 949 sessions, 21.5 hours, 215.6 million tokens, the Claude Mythos 5 model, a database of 1.9 billion protein clusters, 3 to 21 repeats per array — come from The Next Web, citing the technical report; Anthropic's blog gives only rounded numbers. The company says human involvement was limited to the initial prompt and the bench work, and that an analysis of this kind would take an expert scientist weeks to months.

The detail that complicates the story comes from the same source: according to The Next Web, when the search was rerun ten more times, no run reproduced the original discovery. In controlled tests, with the sequences supplied directly, the four most capable Claude models recognized the array in at least 90% of attempts, but only in 32% when they had to get there by working with files and tools. On the institutional side, Anthropic says its biology team — set up in spring 2026, with a lab in the Bay Area — works only at biosafety levels BSL-1 and BSL-2, does not handle human pathogens, and that all lab work is carried out by human scientists. “Eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment, with appropriate safeguards in place, but we aren't doing that today,” the company's chief executive, Dario Amodei, said, as quoted by TechCrunch.

Outside assessment, for now, amounts to a single line. “The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation,” said Feng Zhang of MIT and the Broad Institute: intriguing and worth a closer look, which is exactly as much as can be said about a preprint that hasn't been peer-reviewed and sits on the website of the company that signs it. For a year now, AI labs have been saying their agents will speed up scientific discovery; this is one of the first cases in which one of those companies has brought forward a new biological result, found by agents and partly verified by its own lab. What it is worth will be decided by the scientific community, not by the press release.

What strikes me isn't the 90% in controlled tests, it's the 32% when the models have to use the tools on their own. The gap between those two numbers is the whole craft: not understanding a sequence, but knowing how to find it while working with files and tools. And the fact that none of the ten reruns found the original discovery again doesn't diminish it — but it does say we are still in territory where a real hit and a lucky shot can only be told apart afterwards, and only thanks to someone in a lab coat who goes and checks at the bench.

— Olya