Claude's enzyme find began with five examples from humans

Roughly 950 Claude agents spent 21 hours and 210 million tokens sorting more than 200,000 enzymes down to 20 reports, and one of those reports pointed to a CRISPR-like system. In a September 27 piece, The New York Times asks whether the find, which came out of Anthropic's new biology lab, really counts as a discovery the AI made on its own.

At a glance

• Anthropic's biology lab reported enzymes that could be useful in biotechnology. At the centre is a CRISPR-like system that Claude flagged while scanning DNA databases for reverse transcriptases, enzymes that copy RNA into DNA.
• The search ran about 950 agents for 21 hours on 210 million tokens. They collected over 200,000 reverse transcriptases, narrowed them to 3,500 new candidate systems, then to 20 human-readable reports.
• The enzyme itself was already known from earlier studies. What Claude reportedly spotted first were an attached array of non-coding DNA and an accessory protein, while humans wrote the prompt and ran the lab.

If you haven't been following, Anthropic says it formed a life sciences research group in the spring of 2026. The goal was to see whether general AI models can "systematize and accelerate" the kind of finds that gave biology restriction enzymes, Taq polymerase and CRISPR, each of which started with a scientist noticing something odd in nature. According to Anthropic, this required building its own lab, with a single team handling everything from training Claude in biology to running experiments.

About 950 agents narrowed 200,000 enzymes to 20 reports in 21 hours

A reverse transcriptase, or RT, is an enzyme that copies RNA into DNA. Anthropic gave Claude a single initial prompt: search a huge database of DNA sequences for interesting RTs. Claude agents then combed through the database, looked into distinct RT families and used their own judgment to choose candidates. According to Anthropic, humans only wrote the initial prompt and did the lab work.

The New York Times adds two details. Before the search began, Anthropic scientists gave the AI five examples of reverse transcriptases. The agents then wrote their own software to search the databases for similar genes.

The scale is what stands out. Roughly 950 agents worked for 21 hours and used 210 million tokens. They gathered more than 200,000 RTs, narrowed them to 3,500 new candidate systems and then picked the 20 most compelling, each written up as a report a person can read. Anthropic says this kind of sorting would otherwise take an expert scientist weeks to months.

The enzyme was already known, but its repeat array and mystery protein were not

The RT that Claude flagged comes from a jumbo phage, a large virus that infects bacteria, and earlier scientific studies had already identified it. What Claude reportedly noticed for the first time were the system's other defining features: a nearby array of non-coding DNA sequences and an accessory protein of unknown function.

According to Unite.AI, the first sign came while one agent was reading the raw DNA sequence right next to the RT. The outlet quotes the note the agent logged at that moment:

The DNA next to the RT is spectacular: I can see by eye a tandem repeat array … that’s a CRISPR-like … repeat array?!

According to the same report, the agent then counted the repeats, measured the spacing between them, compared the layout with known RT systems and searched the scientific literature. Only after that did it file a report for humans to review.

Humans wrote the prompt and did all the lab work

According to TechCrunch, Anthropic confirmed its wet lab existed only the week before the enzyme announcement. The lab is in the Bay Area and does only lower-level biosafety work, BSL-1 and BSL-2, with no pathogens that infect humans. TechCrunch reports that human scientists did all of the lab work and that Anthropic kept humans in the loop, never letting Claude run loose in the lab unsupervised.

The same outlet reports that Anthropic CEO Dario Amodei acknowledged on X that the discovery built on other people's work. He said a Stanford team had previously found a system "that is in some ways similar to the one Claude found." He still described the new finding as made "mostly, though not entirely, by Claude."

Reactions are mixed. Business Standard reports that some scientists were quick to cast doubt on the achievement, though the coverage doesn't say which step of the work they question. Anthropic, for its part, quotes CRISPR pioneer Feng Zhang of MIT and the Broad Institute. He reviewed the pre-print and called the RNA-repeat arrays found next to reverse transcriptases "genuinely intriguing" and worth further investigation.

Why does a repeat array next to an enzyme catch a biologist's eye?

In bacteria, CRISPR works as an immune memory. It is a stretch of DNA where short identical repeats alternate with spacers, which are snippets copied from past invaders. Genes that sit right next to such an array often work together with it. So a repeat array beside an RT suggests the enzyme may be part of a larger system rather than acting alone.

The search itself works like a funnel. Hundreds of agents do the broad, tedious part in parallel: they gather RTs, sort them into families and keep the new candidates. Only the 20 strongest reach people, as reports they can read. Picture a crew sorting a warehouse of unlabelled boxes and leaving a one-page shortlist on the scientist's desk.

The coverage we have doesn't say how the system Claude found differs from the earlier Stanford one, or what the lab work actually showed about the array and the protein. In our view, the five seed examples and Anthropic's "single initial prompt" framing don't sit well together: a search that starts from enzymes chosen by humans is a weaker claim to autonomy than the phrase suggests.

What the pre-print still has to answer

The work is still a pre-print, and no date for journal publication has been given. Zhang said the repeat arrays merit further investigation, and nobody has said yet what the accessory protein does. The key open question is how much the new system differs from the one Stanford described earlier, which Amodei mentioned but the coverage hasn't spelled out.