Claude got 21 hours and 950 agents – found a CRISPR-like system

Claude got 21 hours and 950 agents – found a CRISPR-like system

Anthropic claims the company's Bay Area wet lab has delivered its first result: a previously unknown enzyme system in bacteriophage DNA, found "mostly, though not entirely, by Claude," according to CEO Dario Amodei. But there is an important caveat: until experts have replicated the finding, it is only the company's own claim — and the biology behind it is unexplained.

What Anthropic claims

On Wednesday, September 23, 2026, Anthropic announced that the Claude model had helped discover a new enzyme system with "properties reminiscent of CRISPR," as the company describes it (TechCrunch, NDTV). The company presents it as the first result from its biology research program — just five days after the laboratory was publicly confirmed.

The finding is named ART, "array associated reverse transcriptases." The system is found mainly in bacteriophages — the viruses that infect bacteria — and combines three elements: a reverse transcriptase gene, a partner gene, and a long array of evenly spaced DNA repeats (Cryptobriefing). It is the repeat pattern that has drawn attention: it is structurally similar to the arrays that make CRISPR systems programmable. But Anthropic itself has cautioned that ART's function is still unknown, and structural similarity to CRISPR is not the same as gene-editing capability.

The search: 950 agents, 21 hours, 210 million tokens

The search reportedly took place in two phases. First, Anthropic deployed roughly 950 Claude agents that used 210 million tokens to search through biological databases. The agents collected over 200,000 reverse transcriptases, identified around 3,500 candidates, and narrowed the field to 20 (TechCrunch, Cryptobriefing). The analysis work itself reportedly took Claude 21 hours of continuous effort.

Note that all figures come from Anthropic itself. None of the available sources provide independent verification of the agent count, the token consumption, or the time spent — Android Headlines recycles TechCrunch and Amodei's own X posts (Android Headlines).

The lab: humans at the bench, for now

The wet lab was confirmed on September 18 by Eric Kauderer-Abrams, Anthropic's lead for life sciences, in a Reuters interview (HuffPost). The lab was established earlier this year — Anthropic would not say exactly how many months it has been open — and looks like a typical molecular biology laboratory, according to the company.

Anthropic emphasizes that all physical work is performed by human scientists:

"Our laboratory, located in the Bay Area, looks like a typical molecular biology laboratory. We conduct research that involves only the lower biosafety levels (BSL-1 and BSL-2), and we do not handle pathogens that can infect humans. All laboratory work is performed by human scientists," the company wrote in a statement quoted by TechCrunch.

Kauderer-Abrams also defended the investment in physical laboratory work: "We believe that to do biology, the final test is still — and will be for a while — in real lab work," he said. A spokesperson clarified that the lab is not specifically for drug development.

According to Reuters sources (two anonymous people familiar with the matter), Anthropic wants to let Claude direct robotic laboratory units with limited human intervention. But the company's spokesperson says human oversight and participation are considered essential for safety. This is thus reporting based on anonymous sources, not confirmed company policy.

Key caveats

Amodei himself acknowledges that the finding builds on the work of others, and that a Stanford team previously found a system "that in some ways resembles what Claude found," he wrote on X (TechCrunch). The phrasing "mostly, though not entirely, by Claude" is therefore the company's own framing — not an independent assessment of how much of the work can be attributed to the AI model.

The work has been released as a preprint, not peer-reviewed. Anthropic says further experiments are underway, but in the meantime ART's biological function is unknown, and it is up to the research community to validate the finding (Cryptobriefing). Even if the CRISPR-like structure is interesting, it says nothing about whether ART could be used as a gene-editing tool.

Part of a larger buildout

The announcement comes amid a rapid expansion of Anthropic's life-sciences effort. The company has launched Claude Science, a research workbench connected to more than 60 scientific databases with tools for genomics, proteomics, and structural biology. Novartis chief Vas Narasimhan has joined the board, and Anthropic has acquired Coefficient Bio for roughly $400 million in stock — a deal the company confirms is intended to strengthen its drug development capabilities, according to CNBC. In addition come partnerships with Novo Nordisk, Bristol Myers Squibb, and Genentech (Quartz/Yahoo Finance).

The questions that remain

The most obvious one is whether the finding holds: it remains to be seen whether researchers outside Anthropic can replicate the ART observations and determine what the system actually does. It is also unclear how much of the intelligent work Claude actually contributed, given Stanford's earlier similar finding and the underlying reverse transcriptase's previous appearance in research.

The safety side is also worth noting. Anthropic is investing in physical biology work, and per Reuters sources also wants automation, at a time marked by its own researchers' warnings about extinction risk and Amodei's earlier concerns about AI-enabled bioterrorism. The company argues that human oversight makes it safe; but the tension between ambition and warning remains unresolved.

In the meantime, Anthropic has delivered what it claims is a new, unusual biological structure — and an example of how AI agent work can compress search time in biological databases to hours. Whether it is a breakthrough, or an interesting candidate finding that needs months of validation, the next few months will show.