Anthropic says 950 AI agents searched massive DNA data and found a new phage enzyme system called ART
Anthropic said Claude scanned more than 200,000 reverse transcriptases and identified a previously uncharacterized, CRISPR-like enzyme system in bacteriophages called array-associated reverse transcriptase, or ART, showing how AI agents could help mine massive DNA datasets for new biological tools.
The search used roughly 950 AI agents over 21 hours and about 210 million tokens, according to Anthropic, producing about 3,500 candidate systems and narrowing them to 20 for detailed human review.
Researchers said ART has not been shown to work like CRISPR and its biological role remains unknown, but early lab work found its repeat array is expressed as multiple distinct short RNAs, a feature that makes the system notable for gene-editing researchers.
The finding matters for U.S. biotech and AI companies because Anthropic is pushing beyond chatbots into life sciences, building a Bay Area molecular biology lab where human scientists test computational leads generated by Claude.
Anthropic said Claude has found a previously uncharacterized enzyme system in bacteriophages after searching more than 200,000 reverse transcriptases. For biotech researchers and AI investors, the claim is an early but tangible example of a large language model being used for discovery work, not just writing code or summarizing papers. Anthropic's life sciences team said Claude identified what it calls array-associated reverse transcriptase, or ART, a molecular system that appears to pair a reverse transcriptase with a neighboring partner gene and a long stretch of evenly spaced DNA repeats. The company says the system resembles CRISPR in structure, though not yet in proven function, which is why the result matters more as a signal about AI-assisted biology than as an immediate gene-editing breakthrough.
The company described the work as part of a broader genome-mining effort aimed at finding unusual biological systems buried in huge DNA datasets. According to Anthropic, roughly 950 agents spent 21 hours combing through genomic data, using about 210 million tokens. Those agents collected more than 200,000 reverse transcriptases, surfaced about 3,500 candidate systems and narrowed the list to 20 candidates that merited detailed reports for human researchers. That workflow is worth paying attention to. In software, the current AI pitch is speed. In biology, the harder problem is search, specifically how to sift through enormous sequence libraries without drowning scientists in false positives. Anthropic's framing is that Claude can serve as a research copilot for that bottleneck, doing the comparative work that can take experts weeks or months, then handing a much shorter list to humans for validation.
One Claude agent, the company said, homed in on an unusual reverse transcriptase family and then examined the surrounding DNA sequence. It spotted a tandem repeat array next to the reverse transcriptase gene, counted the repeats, measured their spacing, compared the arrangement with known reverse transcriptase systems and searched the scientific literature for prior descriptions. Anthropic said that sequence pattern did not appear to have been recognized previously. The reverse transcriptase itself was not new. Previous studies had already identified that enzyme in a jumbo phage. What Anthropic says Claude helped flag was the broader architecture around it, including the non-coding repeat array and an additional protein whose function remains unknown.
Early lab work, conducted by human scientists rather than the model, found that the ART repeat array is expressed as multiple distinct short RNAs, according to Anthropic. That detail is one reason the system is drawing attention. CRISPR arrays also produce RNAs that make CRISPR-Cas systems programmable, even if Anthropic is being careful not to claim ART does anything similar yet.
Dario Amodei, Anthropic's chief executive, wrote on X on September 23, 2026, that Claude had led the discovery of a molecular machine the company suspects could represent a new gene-editing mechanism. He also said its precise function, biotechnological utility, if any, and significance are not yet clear. That caveat is necessary because biology often produces computational leads that do not turn into practical tools. Still, the company is clearly trying to establish a more ambitious role for frontier AI in science. Anthropic said it has set up a Bay Area molecular biology lab to test candidates generated through these computational searches. Human scientists run the experiments, while Claude is used to search genomic datasets, generate hypotheses, analyze promising systems and help interpret results.
The strategic context here is bigger than one enzyme family. U.S. AI companies have spent the past two years trying to prove that their models can do more than chat, summarize and code. Drug discovery startups have made similar claims for years, but most have focused on proteins, chemistry and target identification. Anthropic is pitching something slightly different, using language-model-style agents to mine raw genomic information for entirely new biological systems that conventional workflows may have missed. That puts the company in a race for credibility as much as capability. The market has heard a lot of "AI for science" promises. What investors and researchers want now are specific examples with named mechanisms, laboratory follow-up and outside validation. Anthropic has at least some of that here. After reviewing the work, Feng Zhang, a CRISPR genome-editing pioneer and professor at MIT and the Broad Institute, said the finding warrants further investigation and that the RNA-repeat arrays associated with reverse transcriptases are genuinely intriguing.
His phrasing matters. It is supportive, but it is not a declaration that Anthropic has found the next CRISPR. The current evidence is architectural and expression-based: a repeat array, a reverse transcriptase, a partner protein and short RNAs produced from the array. Researchers still do not know what the system actually does inside phages or whether it can be turned into a useful biotechnology platform.
The background
There is a good reason this story lands now. CRISPR itself came from painstaking work on strange repeat sequences in microbes that initially looked obscure and disconnected from practical medicine. Since then, scientists have kept finding adjacent or analogous systems that expand the molecular toolbox for editing, copying or inserting genetic material. Anthropic's ART result fits squarely into that hunt, except the first pass was done by AI agents rather than graduate students or postdocs manually trawling databases.
For a U.S. audience, the Bay Area angle also matters. Anthropic is one of a handful of AI labs with the capital and talent to pair frontier models with wet-lab validation. If that model works, it could pull top-tier AI companies into direct competition, or partnership, with biotech platforms that have traditionally owned this part of the workflow. It could also make large genomic databases more valuable, since the bottleneck shifts from data collection to interpretation and experimental follow-through.
There is also a quieter point here that other coverage may miss. Claude did not invent a new enzyme out of thin air. It surfaced a pattern in existing sequence data that humans had not connected into a system. That distinction is important because it sets a more believable near-term use case for AI in science: pattern discovery, ranking and hypothesis generation inside messy, high-volume archives where the data already exists but attention is scarce.
What's next
The immediate question is simple: what does ART do? Anthropic said experiments are underway to determine the function of the reverse transcriptase, the accessory protein and the repeat-derived RNAs, and whether the system's unusual architecture has any biotechnology use. Until those results arrive, the headline is less "new CRISPR rival" than "Claude scans more than 200,000 reverse transcriptases and uncovers a CRISPR-like system," an early proof point that AI agents may be able to find worthwhile biology before humans know where to look.
Frequently asked questions
What did Claude discover after scanning 200,000 enzymes?
Anthropic said Claude analyzed more than 200,000 reverse transcriptases and identified a previously uncharacterized CRISPR-like system in bacteriophages. The system is called array-associated reverse transcriptase, or ART, and may expand the set of biological tools researchers can study.
What is ART in this Claude enzyme discovery?
ART stands for array-associated reverse transcriptase, a newly described enzyme system found in phage DNA. Anthropic characterized it as CRISPR-like, meaning it may share useful properties with programmable defense or editing systems studied in microbes.
How did Anthropic use AI agents to search DNA data?
According to Anthropic, the company used about 950 AI agents over roughly 21 hours to examine a huge biological dataset. The run reportedly consumed about 210 million tokens, showing how AI agents can help researchers sift through sequence data at scale.
Why is a CRISPR-like system in bacteriophages important?
Phages are viruses that infect bacteria, and their DNA can contain unusual molecular machinery with research or biotech value. Finding a CRISPR-like system in phages could give scientists new ways to understand microbial defense, genome engineering, or molecular detection.
Does this mean Claude invented a new gene-editing tool?
No, the report says Claude helped identify a previously uncharacterized enzyme system, not that it created a finished gene-editing product. The finding is an early discovery step that could guide future lab validation and tool development.
What does this Claude biology result mean for AI in science?
It suggests AI agents can do more than summarize papers by actively searching massive scientific datasets for overlooked patterns. In practical terms, that could speed up early-stage discovery in fields like genomics, drug research, and synthetic biology.
