Anthropic says Claude found a new CRISPR-like enzyme system

What Claude found

ART emerged from a search for unusual reverse transcriptases across a large DNA-sequence database. Reverse transcriptases copy RNA into DNA. In this case, Claude noticed one surrounded by a tandem array of repeated sequences and an additional protein whose function is not known.

The system appears mainly in bacteriophages, viruses that infect bacteria. It contains:

The architecture resembles a CRISPR array, which stores a collection of different RNA sequences and helps make CRISPR-Cas systems programmable. Early experiments suggest that the ART array is also transcribed into a set of short RNAs, pointing to a potentially similar mechanism.

That is where the comparison currently stops. Anthropic has not established what ART does. Additional experiments are needed to determine how the system works and whether it can be used as a biological tool.

Claude’s path to the finding looked less like a single clever answer and more like a large-scale filtering operation. The agents:

Feng Zhang, a CRISPR genome-editing pioneer and professor at MIT and the Broad Institute, called the work an interesting example of AI agents contributing to biological discovery. He said that arrays of RNA repeats associated with reverse transcriptases deserve further study and expressed hope that the work would encourage more scientists to use AI in their research.

Anthropic disclosed the result before the system’s main function was known. The company says the early release is intended to demonstrate Claude’s capabilities and give the scientific community a view of work still in progress. It has also published a preprint and a technical report with further details.

The search was designed as a funnel

Anthropic gave Claude a prompt to find interesting new examples of reverse transcriptases in a huge DNA-sequence database. Human involvement was limited to that initial prompt and the laboratory work. The agents examined individual reverse-transcriptase families and selected candidates for further analysis.

They began with more than 200,000 reverse transcriptases, identified 3,500 candidate systems and narrowed those to 20 promising examples. For an experienced scientist, producing this kind of analysis can take weeks to months.

The broader workflow is built around repeated elimination:

The scale changes what counts as useful output. A single campaign can produce several hundred to several thousand candidate reports, so Anthropic is also studying why some hypotheses merit testing while others are set aside. Those findings are fed back into Claude’s instructions, helping the model reproduce the research team’s preferences.

People still run the experiments

Anthropic formed the research group in spring 2026 to test whether general-purpose AI models could organize and accelerate this kind of biological discovery. The group combines scientists who have studied unusual proteins with computational methods for reading DNA, tracing sequence evolution and selecting biological systems for further work.

Its members previously worked on:

The group belongs to Anthropic’s natural sciences division, which also includes teams working on drug development and teaching Claude biology and chemistry.

The laboratory is in the San Francisco Bay Area and operates like a conventional molecular biology lab. Research is limited to BSL-1 and BSL-2, and the scientists do not work with pathogens capable of infecting humans.

All laboratory experiments are performed by people. Anthropic has tested ways to accelerate lab work with AI, including initiatives such as Model Hardware Standard, but says that approach is less suitable for flexible and unpredictable molecular-biology workflows.

Researchers use Claude Science and Claude Code, the same tools available to other scientists. In some cases, they use an internal system that coordinates many Claude sessions working in parallel. Once a candidate passes the team’s review, people produce the protein in standard laboratory strains and study its biochemical and structural properties. Claude helps interpret the resulting data.

The real claim is about research throughput

The historical examples explain why this kind of search matters. Restriction enzymes were found in bacterial immune systems and later became tools for cutting DNA and moving genes between organisms. Taq polymerase, discovered in a Yellowstone hot-spring bacterium, became the basis of PCR. CRISPR began as an unusual repeating DNA sequence in bacteria and now underpins gene-editing medicines.

In each case, an initially obscure molecular pattern became useful only after researchers worked out its function and learned how to control it. ART has not reached that stage. Its value today is as a candidate system and as a demonstration of a search process.

My read is that the more important result is not the name ART or the resemblance to CRISPR. It is the division of labor: Claude searches a space too large for a person to inspect manually, compresses thousands of possibilities into readable reports and proposes one anomaly for experimental attention. Humans still decide what to test and perform every experiment.

The announcement is notably quiet about the part that would turn discovery into technology: what ART actually does. The initial evidence shows a structured repeat array and short RNAs, not a usable biological function. That makes this a credible early-stage discovery story, but not yet a new editing platform.

Anthropic wants to apply the same hypothesis-generation approach to genomics and other fields, and is inviting scientists with research proposals to contact the team. If the model can repeatedly move from unexplained sequence to experimentally validated mechanism, the laboratory becomes more than a showcase for Claude: it becomes a way to expand the number of biological ideas humans can afford to test.