Research

Anthropic says 950 Claude agents flagged a new CRISPR-like enzyme system in 21 hours

In a company-reported result that has not yet been peer-reviewed, AI agents sifted more than 200,000 genes and pointed scientists to a little-studied system inside viruses that infect bacteria.

By Chris · News from Sep 23, 2026 · 4 min read · Checked Oct 2, 2026

Most of us use AI to draft emails or answer questions. But a growing number of scientists are testing whether AI can help with something harder: spotting promising leads in vast piles of biological data. On September 23, 2026, Anthropic, the company that makes the Claude assistant, published a write-up describing one such experiment. The claim is eye-catching, so it is worth reading carefully.

A quick biology primer

A few terms will make the story much easier to follow:

• Gene: a stretch of DNA that carries the instructions for making something in a living cell.
• Enzyme: a molecule that carries out a specific job in a cell, such as cutting or copying genetic material.
• RNA: a close chemical cousin of DNA that cells use to carry and act on genetic instructions.
• Reverse transcriptase: a type of enzyme that copies RNA back into DNA.
• Bacteriophage: a virus that infects bacteria, often just called a phage.
• CRISPR: a defense system bacteria use against viruses, which scientists adapted into a famous tool for editing genes. It is programmable: a short piece of RNA guides it to a precise target.

What Anthropic says happened

According to Anthropic, about 950 Claude agents ran for 21 hours. An agent here means a copy of the AI working on a task by itself, step by step, rather than answering a single question. Together the agents processed about 210 million tokens, the small chunks of text an AI reads and writes, roughly a word or part of a word each.

Their job was to screen more than 200,000 reverse-transcriptase genes. They narrowed that list down to 20 candidates. Among them, they flagged a previously uncharacterized system found in bacteriophages, which Anthropic calls array-associated reverse transcriptases, or ART.

What made it interesting is that this system produces distinct short pieces of RNA. That hints it might be programmable in a way that resembles CRISPR. Anthropic says lab work then confirmed the initial findings. But it also says plainly that the system's actual function is still unknown.

Why the scale matters

The striking part is not a single clever insight. It is the size and speed of the search. Running hundreds of agents at once meant that a screen of more than 200,000 genes finished in under a day, with lab work afterwards to check what they found.

That is the role this experiment points to: AI as a tireless research assistant that sorts a huge haystack into a small pile worth a closer look. It is not AI replacing the scientists who design the experiments and confirm the results.

How to read a company-reported result

This finding comes from Anthropic's own write-up and an accompanying preprint. A preprint is a scientific paper shared publicly before peer review, the process in which independent experts check the methods and conclusions before a journal publishes it. The work has not yet been peer-reviewed, so it is best treated as an early finding.

Independent coverage has kept that framing. Smithsonian magazine headlined its report as what "Anthropic says" its AI found, and noted that the preprint has not been peer-reviewed and that Anthropic's scientists do not yet know what the system does.

One outside voice is on record. According to Anthropic's write-up, Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, reviewed the preprint and called it "an exciting example" of AI agents contributing to biology.

It is also fair to remember that any company has good reasons to publicize wins for its own product. That applies to every AI lab, not just this one, and it is the reason independent review matters.

What we don't know yet

• What the ART system actually does. Anthropic says its function is still unknown.
• Whether peer review will hold up the findings. The preprint has not yet been through independent review.
• Whether other teams can reproduce the result. No independent replication has been reported.
• Whether it will ever become a useful tool, as CRISPR did. A hint of programmability is not the same as a working gene-editing method.
• How much the run cost. Anthropic did not state it.

Sources

• Anthropic: Claude discovers a novel enzyme system, Sep 23, 2026 (company write-up)
• Smithsonian: Anthropic says its AI discovered a new enzyme system that resembles CRISPR