Anthropic’s AI discovers new enzyme system similar to CRISPR

Anthropic has announced that its AI model, Claude, has independently discovered a new enzyme system called ART, which shares similarities with CRISPR—a gene-editing technology that revolutionized biology over the past decade. The findings were published in a preprint by the company, which has yet to undergo peer review.

The discovery of the ART system

The newly discovered system was named ART, short for “array-associated reverse transcriptases.” It is primarily found in bacteriophages—viruses that infect bacteria—and consists of three components: an enzyme that converts RNA into DNA, a neighboring gene with an unknown function, and a sequence of evenly spaced DNA repeats, similar to those found in CRISPR systems.

According to Anthropic, the function of ART remains unknown, but this combination of characteristics had previously only been observed in a few other systems, all of which were programmable and capable of DNA operations such as cutting, copying, and pasting.

Feng Zhang, a pioneer in CRISPR gene editing and a professor at MIT and the Broad Institute, reviewed the preprint at the company’s request. “This is a fascinating example of how AI agents can contribute to biological discovery. The identification of RNA repeat arrays associated with reverse transcriptases is genuinely intriguing and warrants further investigation,” he said.

How Claude discovered the system

Anthropic instructed Claude to search a DNA sequence database for interesting examples of RNA-copying enzymes. Claude’s agents then analyzed over 200,000 sequences, identified 3,500 potential new systems, and narrowed it down to 20 promising cases, which were documented in reports written by the model itself.

The search took 21 hours in total, involved around 950 agents working in parallel, and consumed 210 million tokens—a unit used to measure AI model processing.

During this search, one agent noticed an unusual pattern of DNA repeats near the enzyme gene. “It’s incredible: I can see a series of repeats with my own eyes… is this a CRISPR-like repeat array?!” the agent wrote, according to a passage shared by Anthropic.

The system then mimicked how a scientist would approach a potential discovery: it counted the repeats, measured the spacing between them, compared the pattern to known systems, and searched scientific literature to see if anything similar had been documented before. Only then did it generate a report for human review.

Reproducibility remains limited

Anthropic noted that the result is not easily reproducible. After the initial discovery, the company repeated the search process ten more times using the same initial prompt. In none of these attempts did the agents revisit the specific DNA segment containing the discovery, and all failed to find it again.

In more controlled tests, where researchers provided the model with the relevant DNA segment instead of letting it search an entire database, the company’s four most advanced models identified the repeat pattern in at least 90% of attempts.

However, when the models had to use files and tools independently—as in the original discovery—the success rate dropped to as low as 32%, according to the specialized website Daily.dev. This means the result largely depends on the unpredictable behavior of AI agents rather than a guaranteed discovery method.

The enzyme itself was not new; previous studies had identified it in a type of bacterial virus known as a giant phage. According to Anthropic, Claude’s breakthrough was recognizing the accompanying characteristics—the repeat array and the neighboring gene—which no one had previously linked to it.

Part of a larger investment in biology

This announcement marks the first result from Anthropic’s new laboratory, confirmed on September 18 in San Francisco Bay. The facility operates at the lowest biosafety levels and does not handle pathogens capable of infecting humans, according to the company. All lab work is conducted by human scientists, with Claude assisting in data interpretation.

This initiative is part of Anthropic’s broader recent efforts in the field. In April, the company acquired biotechnology startup Coefficient Bio for $400 million. In June, it launched Claude Science, a platform integrating over 60 scientific databases. Its main competitor, OpenAI, released GPT-Rosalind—a model dedicated to life science research—in the same month.

In a post on X, Anthropic CEO Dario Amodei cautiously commented on the result. “Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud of,” he wrote.

Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism. Its precise function, biotechnological utility (if any), or level of significance is not yet clear, but at minimum it is work I would have been proud… https://t.co/hyzhzirNJe

— Dario Amodei (@DarioAmodei) September 23, 2026

Frequently asked questions

What did Claude discover in biology?

According to Anthropic, Claude identified an enzyme system called ART, short for “array-associated reverse transcriptases.” The system combines an enzyme that converts RNA into DNA, a neighboring gene with an unknown function, and DNA repeat sequences similar to those found in CRISPR systems.

Is the ART system a new version of CRISPR?

No. Anthropic states that ART shares characteristics with CRISPR systems, but its function remains unknown. The company says the specific combination of enzymes and DNA repeats has never been described before.

How did Claude find the ART system?

Anthropic tasked Claude with searching DNA databases for interesting enzyme sequences. AI agents analyzed over 200,000 sequences, selected 3,500 candidates, and narrowed it down to 20 systems deemed promising for further analysis.

How long did it take Claude to make the discovery?

According to Anthropic, the search took 21 hours, involved around 950 AI agents working in parallel, and consumed approximately 210 million tokens—a unit used to measure AI model processing.

Has Claude’s discovery been confirmed by science yet?

Not yet. The findings were published in a preprint, a scientific article that has not yet undergone peer review. Anthropic says the function of the ART system and its biological significance still need to be investigated.

Was Claude able to repeat its own discovery?

Not consistently. Anthropic repeated the same search ten times, and in none of them did the agents find the specific DNA segment associated with the original discovery again. In more controlled tests, where the model was given the relevant segment, the pattern was identified in at least 90% of attempts.

What is Anthropic doing in the field of biology?

Anthropic has established a biological research laboratory in San Francisco Bay and is expanding its scientific activities. The company acquired startup Coefficient Bio for $400 million and launched Claude Science, a platform providing access to scientific databases.

Who reviewed Claude’s discovery?

The preprint was reviewed by Feng Zhang, a professor at MIT and the Broad Institute and one of the researchers involved in developing CRISPR-based gene editing technologies. He described the work as an interesting example of how AI agents can contribute to biological discoveries.