Claude read a phage genome, noticed a repeat array, and filed a report for humans

Anthropic says Claude discovered something new in biology, and its announcement, published on September 23, is careful about what that means. The machine found a pattern in raw viral DNA that no one had described and wrote a report about it. Humans then did the science: they ran the experiments, named the system, and released a preprint that states plainly that its function is unknown. The distance between those two halves is where the story is.

What Anthropic announced

The company has formed a life-sciences research group and built its own molecular-biology lab in the Bay Area, and the September 23 post is the first public result from it. The lab runs only BSL-1 and BSL-2 work and handles no pathogens that can infect humans, and every experiment in it is performed by human scientists. Anthropic states that boundary itself rather than leaving it to inference.

The result is a system, not a single enzyme. Anthropic says Claude was given a single prompt: search a large DNA database for interesting new examples of reverse transcriptases, the enzymes that copy RNA into DNA. Roughly 950 agent sessions ran for 21 hours and spent 210 million tokens. One agent flagged a repeating pattern of DNA sequences sitting next to the gene for an odd-looking reverse transcriptase. After laboratory work by the company's scientists, Anthropic named the system array-associated reverse transcriptases, or ART: the RT itself, a partner gene beside it, and a long array of evenly spaced DNA repeats.

What the machine actually did

The post describes a pipeline that reads better as a chain of custody than as a discovery. Humans wrote the prompt. The agents gathered more than 200,000 reverse transcriptases, picked out 3,500 new candidate systems, and narrowed those to the 20 most compelling, each with a human-readable report proposing a function and the evidence for it. Anthropic notes that this kind of analysis can take an expert scientist weeks to months.

The behaviour worth reading closely is a side observation. The agent's own words, quoted in the post, are not the language of a settled hypothesis:

"[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array ... that's a CRISPR-like ... repeat array?!"

What Claude appears to be first to notice is not the enzyme. The reverse transcriptase itself had already been identified in earlier studies of a jumbo phage. The novelty is the array of non-coding DNA sequences beside it and an additional accessory protein of unknown function.

Why the CRISPR comparison carries the story

The analogy earns its place. A CRISPR array holds the bank of RNA sequences that makes CRISPR-Cas systems programmable, and the ART array is laid out the same way. Anthropic's first experiments show the ART array is also expressed as a set of distinct short RNAs, suggesting something analogous may be at play. The preprint puts numbers on both the resemblance and the limits. ART arrays span 0.3 to 4.1 kilobases, built from 3 to 21 copies of a repeat 15 to 49 nucleotides long with a palindromic core. The spacers between repeats run 120 to 220 nucleotides, against roughly 30 in a CRISPR array, and repeats are conserved within a clade but not across clades. No cas genes sit near any ART locus, which is why the authors call the array a new type of non-coding repeat element rather than a CRISPR relative.

That distinction matters commercially as much as it does scientifically. CRISPR is a tool because its RNA bank is programmable and its effector cuts where the RNA points. For ART, that is a hypothesis with one piece of supporting evidence: the array is transcribed.

Where the vendor's account ends

Anthropic is simultaneously the model vendor, the lab, the benchmark builder and the publisher of the press release, four roles that normally sit with four different parties. The most useful details in its own preprint are the measure of what has actually been established.

The work is a preprint and not peer-reviewed, and no group outside the company has replicated it. More consequentially, Anthropic reran the campaign ten more times with the same harness and brief, and, in the preprint's own words, the array was missed in every rerun. The authors attribute that to the size of the search space and the non-deterministic behaviour of the harness. When they switched to a fixed-input benchmark and handed models the ART loci directly, the most capable Claude models described the array in at least 90% of attempts with the sequence in context, but in as few as 32% when the same loci arrived as files with tools. Reading at least 200 contiguous nucleotides lifted recognition by 16 to 32 percentage points.

The benchmark also had a Claude in the judge's chair: grading was performed by a language-model judge, Mythos 5, the same model family that ran the original campaign. Anthropic is explicit that it treats the awkwardness as data. With hundreds to thousands of candidate reports from a single campaign, the company's own index of how much of its research Claude leads is the house precedent for this kind of self-measurement, and the post says Anthropic studies which proposals its scientists judge worth testing, then feeds that judgement back into Claude's instructions.

How it landed

The reception was small and sceptical rather than celebratory. In the Hacker News thread on the post, the earliest reply to draw agreement worried about provenance: "This is great, but I can't help but wonder if we're going to have another post next week with a lab complaining that they were about to publish this same finding, and they had Claude proofread their paper, and whoops how'd that get into Anthropic's training data?" The answer from the thread below was blunter. The reverse transcriptase had been described before; what Claude commented on was the repeat.

Not everyone read it as theatre. Feng Zhang, one of the pioneers of CRISPR genome editing and a professor at MIT and the Broad Institute, reviewed the preprint and called the identification of RNA-repeat arrays associated with reverse transcriptases "genuinely intriguing and merits further investigation".

The two doors

Whatever ART turns out to be, the organisation that found it is also the organisation deciding who gets to work with the tools such a system might become. Six days before the enzyme post, Anthropic opened its Life Sciences Verification Program, which gives verified life-science professionals access to its Mythos, Opus and Sonnet models "with a refined set of safeguards more permissive for biology-related work". The program's second tier, High-risk Use, "removes all safeguards that block life sciences requests" for a single research project rather than a whole team. Both tiers sit behind reviews of research credentials, security standards and ethical oversight.

That is a policy argument the company makes in the open, and it can be defensible: vetted researchers get access a general user should not. It also means the same week contains a publicised pipeline for mining genome databases at 210 million tokens per campaign and a program that loosens the biology filters on the model doing the mining.

What would settle it

Four things, in rough order of how much they would move the needle. Peer review of the preprint, which currently carries no external stamp. Independent replication of the ART array by a group that does not work for Anthropic. A function: whether the array really acts as a bank of distinct RNAs, and what the accessory partner protein does with them. And, hardest but most telling, whether another group can reproduce the discovery from a comparable prompt. Anthropic's own reruns suggest the answer is not automatically yes.

Until then, the accurate version of the headline is that a model noticed a pattern in a public database, wrote it up, and handed it to humans who ran every experiment. That is a real result about how research gets done, and a much smaller claim than the one the phrase "Claude discovers a novel enzyme system" carries.