Array-associated reverse transcriptases (ART) are a family of reverse transcriptase (RT) systems in the genomes of jumbo bacteriophages, announced by Anthropic on September 23, 2026 as early results from one of the first research programs of a new life sciences research group and laboratory at the company.[1][2] Anthropic says an autonomous agent running on Claude Mythos 5 noticed the defining feature of these systems, a repeating pattern in non-coding DNA next to the RT gene, while reading raw sequence during a large genome-mining campaign.[1][2] The function of ART was not known at the time of the announcement, and Anthropic said further experiments were under way.[1][2]

The work was released as a company blog post plus a technical report, "Autonomous AI agents discover reverse transcriptases with tandem repeat arrays," credited to Peter H. Yoon, Januka S. Athukoralage, Emmanuel Ameisen, Eric Kauderer-Abrams, Nicholas T. Perry and Matthew G. Durrant of Anthropic, San Francisco.[2] The report is hosted on Anthropic's own content server and had not been peer reviewed.[2][9][12] The reverse transcriptase itself was already in the literature: a 2021 comparative-genomics paper on three Staphylococcus aureus jumbo phages, including the phage MarsHill, recorded a "retron-like reverse transcriptase" in those genomes.[3] Anthropic's claim of novelty is narrower, that Claude "appears to be the first to notice the system's defining features," the associated array of non-coding DNA and an accessory protein of unknown function.[1] The technical report makes the same point about that 2021 paper, noting that it identified the RT and proposed a non-coding RNA upstream of it but described neither the repeats nor the partner gene.[2][3]

Background

Reverse transcriptases copy RNA into DNA, an activity that underlies laboratory methods from cDNA synthesis to genome engineering.[2] In bacteria, RTs are frequently encoded beside a partner protein and a non-coding RNA in the same genomic locus, and many of these systems act in defense against phages.[2] One well-studied example is the retron, where the RT works with a dedicated non-coding RNA and an effector protein; the Anthropic report treats retrons as the closest described relative of ART and places the ART enzymes in a clade next to them on its phylogenetic tree.[2]

Most known bacterial RT families were found by genome mining, that is, by searching sequence databases for uncharacterized genes, noticing the unusual ones and then working out what they do.[1][2] The report argues that automated pipelines sort database hits by features chosen in advance, so anomalies outside those features go unseen unless an expert inspects the data by hand, and that this expert curation does not scale with the growth of sequence databases.[2] The stated motivation for the campaign was to test whether agentic large language models could take over some of that curation, a question that sits in the wider push to apply AI to scientific discovery.[2]

The agent campaign

Anthropic wrote a research brief asking agents to identify novel RT systems on the basis of new partner-gene associations. According to the report, neither that brief nor the task brief under which the array was found mentioned repeats or arrays.[2] Each agent was an instance of Claude Code configured with Claude Mythos 5, coordinated by an in-house harness that split the brief into five stages and assigned each task to a worker agent, with a supervisor agent reviewing plans and results and opening follow-up tasks, a curator agent maintaining a shared knowledge base, and an editor agent reviewing reports.[2] The agents worked in a sandbox with standard bioinformatics software, public sequence and structure databases, literature search and a library of written method guides.[2]

The search space was an in-house database of proteins predicted from public metagenomic and genomic assemblies spanning roughly 11 million biosamples, drawn mainly from Logan, the European Nucleotide Archive, the Joint Genome Institute and NCBI. Clustering at 50% identity yielded 1,939,242,578 clusters, the "1.9 billion protein clusters" cited in Anthropic's summaries.[2]

Figure	Technical report	Blog post
Agent sessions	949	"roughly 950 agents"
Wall-clock time	21.5 hours	21 hours
Tokens	215.6 million	210 million
Agent-hours	76.9	not stated
RT clusters recovered	198,290	"over 200,000 RTs"
Candidate partner families scored	3,564	"3,500 new candidate systems"
Reports filed	19	"20 most compelling candidates"
Tasks	119	not stated

Sources: Anthropic technical report and blog post.[1][2]

The report's token figure is the sum of uncached input tokens (11.3 million), output tokens (14.9 million) and tokens written to the prompt cache (189.5 million); tokens read from the cache were excluded.[2] Of the 119 tasks, 98 were follow-ups that the agents themselves opened.[2] Seventeen candidate partner families were promoted for detailed study, of which only three were confirmed as previously unreported RT associations; workers separately flagged three new RT lineages, one of which became ART.[2] The 19 reports were ranked in a tournament of pairwise comparisons scored by a judge model, in which the ART report placed third.[2]

How the array was noticed

The path to ART was indirect. Agents first selected a candidate because a particular RT lineage sat next to a jumbo-phage RNA polymerase subunit gene. A worker rejected that association as spurious but queued a follow-up, reasoning that a free-standing retron-like RT inside a jumbo phage was itself worth reporting.[2] Because retron RTs work with a non-coding RNA encoded upstream, a supervisor asked the next worker to check the regions upstream of these RTs. Those RTs had only short upstream regions (median 27 bp), but their closest relatives had long ones (median 940 bp). The worker loaded the relatives' upstream DNA directly into its context and wrote that one flank was "spectacular: I can see by eye a tandem repeat array," asking whether it was "a CRISPR-like or msDNA-like repeat array?!"[2] Anthropic's blog post quotes an abridged version of the same line.[1] The transcript then shows the worker questioning its own novelty claim and listing known systems it might match, before quantifying the pattern and running a literature check.[2] The report states that no repeat-finding tool was run in that session and that no tool call intervened between the sequence retrieval and the recognition.[2]

The ART system

As defined in the report, an ART locus has three parts: the RT, a dedicated partner gene directly downstream, and an array of repeats in the non-coding region upstream.[1][2] Profile searches of Anthropic's database and public genomes identified 95 distinct RT clusters of the family in cultured jumbo phages and predicted viral contigs, 28 of which had a detectable array upstream.[2] The ART enzymes carry the catalytic motif shared by characterized RTs and their polymerase domain superimposes on solved retron and diversity-generating retroelement structures, but they have an unusually long N-terminal extension of about 180 residues where other RTs typically have about 50 or fewer, and that extension shows at most weak similarity to any known sequence family or fold.[2]

The arrays span roughly 0.3 to 4.1 kilobases and hold 3 to 21 copies of a short repeat at near-constant spacing, with repeats 15 to 49 nucleotides long and intervening spacers of 120 to 220 nucleotides. The report's abstract summarizes the arrays as built from units of about 200 nucleotides.[2] The report contrasts this with CRISPR arrays, where near-identical repeats alternate with spacers of about 30 nucleotides that are gained and lost between related strains; in ART, spacers are longer, are retained in order between related phages, and no cas genes occur near any ART locus.[2]

Anthropic re-analyzed published RNA sequencing data from an infection time course of the Staphylococcus jumbo phage SA1, a 2022 study from a group in China, and reported that the array region is highly expressed during infection and resolves into shorter RNA species with reproducible boundaries.[2][4] At 15 minutes after infection the array-derived RNAs accounted for up to 8% of phage RNA, among the most abundant phage transcripts.[2] The team also expressed parts of the SA1 locus in Escherichia coli and recovered similar discrete short RNAs.[2] Three unrelated families of partner protein were found beside ART enzymes, and the report classifies ART systems into three types accordingly.[2]

From these observations the authors propose, as a hypothesis rather than a result, that ART may work like a retron supplied with a bank of different RNAs, so that one enzyme and partner pair could form several complexes differing only in their RNA.[2] They state plainly that they have not shown the RT is active, that the array RNAs are its substrates, that the RT and partner interact, or what the system does for the phage.[2]

Reproducibility and model benchmarks

Anthropic re-ran the same campaign ten more times with the same harness and brief. Nearly every run that completed the census sampled ART loci and two investigated the lineage, but none read the DNA upstream of the RTs, and the array was missed in every rerun; the authors attribute this to the size of the search space and the non-deterministic behavior of the harness.[2] They also note that a locus outside the set of identifiers they searched for would not have been detected by that audit.[2]

To study the recognition step in isolation, the team built a fixed-input benchmark in which a model receives ART sequences at one of five levels of information, from two protein sequences in context up to all 96 loci as files with analysis tools, structures, literature and web access. Each attempt produced a report scored by a judge model against ten curated claims about the system.[2] Seven Claude models were run for 100 attempts at each level. Four models (Opus 5.5, Mythos 5.1, Mythos 5 and Opus 5) scored clearly above the other three (Opus 4.6, Opus 4.8 and Sonnet 5).[2] Given the loci directly in context, the four stronger models described the array in at least 90% of attempts; with files and tools the rate fell as low as 32% (Opus 5 at level 4). The authors traced the drop to models not reading the sequence: with files, 39% of the four models' attempts never read a contiguous stretch of 200 nucleotides or more and so never saw more than about one repeat unit, and reading at least 200 nucleotides raised recognition by 16 to 32 percentage points for each model.[2] Recognition rose with the amount of sequence read into context, rising from 29% to as high as 76% for the four models pooled and as high as 96% for Mythos 5.[2]

Interpretability analysis

The report includes an interpretability section examining what happened inside the model as it read the sequence. The transcript of the discovery session was replayed through a separately hosted copy of the same Mythos 5 checkpoint, and the model's activity at one layer was decomposed with a sparse dictionary, the technique Anthropic has previously described in its dictionary learning and sparse autoencoder work.[2] Twelve candidate signals were selected in advance on synthetic repeats planted in computer-generated sequences, before the real flank was examined; the two shown in the report were chosen from those twelve after the flank had been examined, one automatically labeled as responding to tokens repeating an earlier pattern verbatim and one to letters in repeated sequence motifs and garbled text.[2] Both were silent before the array. One peaked at the third repeat copy and then faded along the array, while the other strengthened by the third copy and stayed steady, with its activity largely confined to the repeat itself. Shuffling each copy in place silenced the first signal on 12 of the 14 copies and the second on all 14.[2] The authors note that neither signal is specific to DNA, since both also fire on repeats planted in random strings of letters and digits.[2]

For comparison, the same region was scored with two genomic language models, Evo 2 and gLM2, whose likelihood profiles also picked out the repeats.[2] The report says the two Mythos 5 signals responded to the array in a similar fashion to these models trained only on biological sequences, and its discussion argues that the benchmarks and the internal signals together suggest a "genomic vision" (recognizing the array by reading raw DNA) that directly enabled the discovery.[2]

Anthropic's life sciences lab

Anthropic says it formed the research group in spring 2026 and that its Bay Area lab "looks like a typical molecular biology lab," working only at biosafety levels 1 and 2 and handling no pathogens that infect humans, with all laboratory work performed by human scientists.[1] Computational work is done in Claude Science and Claude Code, and sometimes with the in-house harness that coordinates parallel sessions.[1] The group sits inside Anthropic's life sciences organization, which also covers drug discovery and training Claude in biology and chemistry.[1] The wet lab had been reported by Reuters and confirmed to TechCrunch on September 18, 2026, five days before the ART announcement. The company declined to say what the lab was working on but said its main focus was fundamental biology rather than drug discovery; its head of life sciences, Eric Kauderer-Abrams, told Reuters that "to do biology, the final test is still, and will be for a while, in real lab work."[8] Anthropic's post also invites outside scientists to propose research questions.[1]

CEO Dario Amodei promoted the result on X on September 23, describing it as "the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism," while adding that "its precise function, biotechnological utility (if any), or level of significance is not yet clear."[10] He wrote that the work was done "mostly, though not entirely, by Claude," with the life sciences team suggesting the research area and carrying out the experiments Claude proposed, and said Anthropic is not today letting Claude run lab equipment autonomously.[10] He also noted that a Stanford team had independently described a different RT system with an associated non-coding array.[10] That work, a preprint from Brian Hie's group posted to bioRxiv on September 23, 2026, reports arrays of non-coding RNAs beside unrelated UG27 RTs found with a purpose-built genome language model; Anthropic's report cites it as evidence that the architecture arose more than once.[2][11]

Reception

Coverage repeated Anthropic's own caution that the system's function is unknown.[5][6][7] Feng Zhang of MIT and the Broad Institute, who is thanked in the report's acknowledgments, along with three others, for reviewing an early copy of the manuscript, was quoted in Anthropic's post: "This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA-repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation. I hope this work encourages more scientists to explore how AI can support their research."[1][2]

Writing in The Conversation, Dimitri Perrin of Queensland University of Technology said the enzyme was already known and that "what is new is the recognition that it may form part of this larger system," concluding that "at this point in time, we can say that ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function."[6] He argued the significant part is autonomy rather than the biology, since none of the individual techniques are new and a purpose-built pipeline could probably have found the same pattern, and that the study does not report how many of the 3,500 candidates or the 20 shortlisted ones were genuinely novel, so "Finding one intriguing result after exploring thousands of possibilities is not the same as showing that the system can reliably recognise discoveries."[6]

Al Jazeera quoted Stanley Qi of Stanford, who called the result "incredibly exciting" and said what stood out was "its ability to recognize an unusual biological pattern that was difficult to detect before," and Kevin Blake of Washington University School of Medicine, who warned that "CRISPR-the-technology is very different from CRISPR in nature" and that "there's nothing to indicate this is a rival to CRISPR-the-technology, or could be developed into any kind of therapeutic or practical application."[5] Gizmodo, which carried both Perrin's and Blake's criticisms, also quoted an X post by Bo Wang, whom it identified as chief artificial intelligence scientist at Toronto's University Health Network, saying the opportunity is "scaling scientific attention."[7]

The Next Web, summarizing the preprint, emphasized how provisional the result is, noting that the team has not shown the enzyme is active or that it acts on the array RNAs, and that ten reruns of the campaign all missed the array.[9]

The Scientist framed the announcement as the public debut of the company's wet lab and noted that the discovery, "made with a team of six named human scientists, also used 210 million tokens," linking it to earlier concerns among academics about the cost of AI tools relative to funding for trainees.[12]

The Hacker News thread on the announcement drew 780 points and about 800 comments.[13] Recurring criticisms there were that presenting the work as something "Claude" did erases the human scientists on the paper, that a company blog post and a self-hosted preprint are weaker than peer review, and that the underlying enzyme was already described so the word "discovery" is doing heavy lifting; other commenters countered that the authors are established researchers in the field and that preprints are normal practice in fast-moving areas.[13]

Open questions

Anthropic's own stated limitations are that ART's primary function is unknown, that the RT has not been shown to be active, that the array RNAs have not been shown to be its substrates, and that no interaction between the RT and its partner protein has been demonstrated.[2] The report adds that harness configuration and model capability strongly affect whether such a discovery happens at all, given that ten reruns of the same campaign missed the array.[2] Independent commentary has added that the architectural resemblance to CRISPR says nothing about function, and that a single success does not establish a reliable discovery rate.[6]

References

Claude discovers a novel enzyme system with CRISPR-like repeats - Anthropic, September 23, 2026
Autonomous AI agents discover reverse transcriptases with tandem repeat arrays - P. H. Yoon, J. S. Athukoralage, E. Ameisen, E. Kauderer-Abrams, N. T. Perry, M. G. Durrant, Anthropic technical report (PDF), 2026
Comparative Genomics of Three Novel Jumbo Bacteriophages Infecting Staphylococcus aureus - A. M. Korn, A. E. Hillhouse, L. Sun, J. J. Gill, Journal of Virology 95(19), September 9, 2021
Interactions between Jumbo Phage SA1 and Staphylococcus: A Global Transcriptomic Analysis - B. Zhang, J. Xu, X. He, Y. Tong, H. Ren, Microorganisms 10(8):1590, August 7, 2022
AI model Claude discovers CRISPR-like enzyme system, Anthropic says - John Power, Al Jazeera, September 24, 2026
An AI model has found a new 'CRISPR-like' biological system. Here's what it means for science - Dimitri Perrin, The Conversation, September 24, 2026
Claude Found a Mysterious CRISPR-Like System, but Anthropic Can't Say What It's Capable of - Matthew Phelan, Gizmodo, September 24, 2026
Anthropic is operating a lab that conducts biology experiments - Julie Bort, TechCrunch, September 18, 2026
Anthropic says Claude found a new enzyme system with CRISPR-like repeats - Ana Maria Constantin, The Next Web, September 23, 2026
Dario Amodei on X: "Today we announced the Claude-led discovery of a molecular machine..." - September 23, 2026
Coevolutionary mining of prokaryotic non-coding elements with a genome language model - D. B. Li et al., bioRxiv preprint, September 23, 2026
Anthropic's Secretive AI-Powered Wet Lab Breaks Cover and Makes First Discovery - RJ Mackenzie, The Scientist, September 24, 2026
Claude discovers a novel enzyme system with CRISPR-like repeats - Hacker News discussion, September 23, 2026

Editing: Array-associated reverse transcriptase (ART)