Anthropic says Claude autonomously spotted a novel enzyme system in bacteriophage DNA, which its biology lab named array-associated reverse transcriptases, or ART
About 950 Claude agents processed 210 million tokens over 21 hours, narrowing 200,000 candidate reverse transcriptases down to 3,500, then to 20 compelling candidates before one flagged the pattern
ART resembles CRISPR arrays and looks programmable for DNA operations, but Anthropic states plainly it does not yet know the system's function
The result is a preprint, not peer-reviewed, and every physical lab experiment was carried out by human scientists, not the model
What Anthropic Actually Announced
On September 23, Anthropic published a result from its biology research lab describing a novel enzyme system found in the DNA of bacteriophages, the viruses that infect bacteria. The lab calls it array-associated reverse transcriptases, shortened to ART. It has three parts: a reverse transcriptase enzyme, a partner gene sitting next to it, and a long array of evenly spaced DNA repeat sequences. That last piece is what makes the comparison to CRISPR obvious to anyone who has followed gene editing at all, since CRISPR systems are built around a similar kind of repeating array.
The underlying reverse transcriptase was first found in a jumbo phage, a category of unusually large bacteriophage, but Anthropic is careful to separate two different claims. The enzyme itself was already known to exist. What Claude identified was the full system, the enzyme plus the neighboring gene plus the repeat array acting together as one unit, a pattern that had not been described before. That distinction matters for judging how big this actually is: this is not "Claude found a new molecule," it is "Claude found a new relationship between molecules that scientists had already been looking at separately."
Anthropic is direct about what remains unknown. The lab's own language is that it does not yet know the system's function. Early experiments show the ART array gets expressed as a set of short RNAs, which hints that something CRISPR-like could be happening, but that is a suggestive early signal, not a confirmed mechanism. This is a preprint, released to share the finding early, and it has not gone through peer review. I want to be precise about that boundary here rather than let the CRISPR comparison run ahead of what has actually been shown.
How the Discovery Actually Happened
The methodology is the part of this story that is easiest to overstate, so I want to walk through the actual numbers rather than the shorthand version. Anthropic's research group, formed in the spring of 2026, set roughly 950 Claude agents loose on a genomic database. Over 21 hours, those agents processed about 210 million tokens of data. That scale alone says this was never a single conversation with a model, it was a distributed search running continuously across hundreds of parallel agent instances.
The funnel matters more than the headline number of agents. The search started by gathering more than 200,000 reverse transcriptases from the database. That set got narrowed down to 3,500 candidates worth closer inspection, then narrowed again to 20 compelling candidates. Somewhere in that final round, one agent noticed an unusual recurring pattern, the repeat array, and flagged it for human review rather than continuing past it as noise. That flagging step is the actual discovery moment inside the whole process, a single agent treating an odd pattern as worth stopping for instead of filtering it out.
Everything after that flag was human work. Anthropic states plainly that the physical lab experiments, the part where a hypothesis from a database search gets tested against real biological material, were carried out by human scientists at the company's Bay Area laboratory. Claude searched, filtered, and flagged. People verified. That division of labor is worth being exact about, because it is the difference between "an AI model did biology" and "an AI model did a very large scale search that human biologists then had to confirm was real." Anthropic's own framing lands closer to the second description, and I think that framing is the honest one.
I keep coming back to the funnel numbers because they are the actual evidence of what changed here, more than any single word in the announcement. Going from 200,000 candidates down to 3,500 is a filter a human team could, in theory, eventually work through by hand given enough time. Going from 3,500 down to 20 compelling candidates is the step that starts to require real judgment about what counts as compelling, not just a mechanical filter on database fields. That is the part of the funnel where I think the agents were doing something closer to expert triage than search, and it is also the part most likely to hide both false negatives, a real pattern discarded too early, and false positives, a coincidence mistaken for a signal. Neither risk is unique to AI-driven search. Human-led genomic screens have the exact same failure modes. What is different here is the sheer volume the funnel could process before a human ever had to look at a single candidate.
Why the CRISPR Comparison Is Doing a Lot of Work
CRISPR is one of the few biology terms that carries instant recognition outside biology itself, which is exactly why every headline about ART reaches for it immediately. The comparison is not baseless, repeat arrays are a real structural similarity, and CRISPR-associated systems are themselves built around reverse transcriptase-adjacent machinery in some known variants. But "resembles CRISPR structurally" and "functions like CRISPR" are different claims, and only the first one has any real support right now.
What CRISPR actually does, at a level worth restating here, is let a cell target and cut specific DNA sequences with precision, which is what made it usable as a gene editing tool once scientists understood the mechanism well enough to redirect it deliberately. ART has the repeat-array structure that made CRISPR recognizable, and Anthropic's early RNA expression data hints at a similar kind of targeting behavior, but nobody has yet shown ART cutting, copying, or pasting anything in a controlled experiment. The "suspected to be programmable" language in Anthropic's own announcement is doing exactly the work that phrase implies: a hypothesis based on structural resemblance, not a demonstrated capability.
I bring this up because I have written before about being careful with early AI-driven research claims before they clear peer review, in the same spirit as how I covered Claude formalizing Fermat's Last Theorem in Lean, a result that was also genuinely impressive and also worth stating precisely rather than inflating. ART deserves the same treatment. A large-scale search that surfaces a real, previously undescribed biological pattern is a significant result on its own. It does not need an unearned "it's the next CRISPR" claim stacked on top to be worth writing about.
What This Says About Where Claude's Research Work Is Heading
I covered Anthropic's own numbers on how much of its internal research now runs through Claude in an earlier piece on Claude handling roughly a quarter of the company's AI research work, and this result reads like a concrete instance of that pattern showing up somewhere outside AI research itself, in a completely different field. Biology is not the domain Claude was primarily trained to reason about, and a 950-agent search across a genomic database is a different kind of task than writing code or drafting a document. Watching that scale of search work land a real, previously unnoticed pattern in a field this far from software is the part of the story I find most interesting, more than the CRISPR comparison itself.
I also noticed how similar the shape of this process is to good engineering search work I recognize from my own domain: cast a wide net, apply a cheap filter to cut the set down by orders of magnitude, apply a more expensive filter to the survivors, and have a human make the final call on the few candidates left. That is not a coincidence. It is the same funnel shape that shows up anywhere a search space is too large for a human to review directly but small enough, once filtered, for expert judgment to take over. Seeing that pattern work in biology instead of software is a genuinely interesting data point about how generally that approach transfers.
It also raises a question I do not think Anthropic has fully answered yet, which is how much of this scales with more compute versus how much depended on this particular database having a findable pattern in it at all. A search that spends 210 million tokens across 21 hours and comes back empty is not a failure exactly, it is a null result, but it also does not make headlines. I would want to see how often this kind of large-scale genomic search comes back with nothing notable before treating a 20-candidates-from-200,000 hit rate as the expected outcome rather than a genuinely lucky one. Anthropic did not publish that base rate, and I think that is the honest gap in an otherwise carefully hedged announcement.
Bottom Line
Claude helped surface a real, previously undescribed enzyme system by searching a database at a scale no human team reviews by hand, then handed the actual verification to human scientists who ran the physical experiments. That is the accurate version of the story: a large filtered search plus human confirmation, not an AI model independently doing biology end to end.
The CRISPR comparison is structurally fair and functionally unproven, and Anthropic's own preprint says as much if you read past the framing. I will treat ART the way I treat any preprint, interesting and worth watching, not settled. What I take from this result is less about gene editing and more about the search pattern itself: narrow a huge space with cheap filters, escalate the survivors, and let a human close the loop. That pattern is showing up everywhere Claude gets pointed at a large enough problem, and biology just became the newest place to watch it work.
Top comments (0)