950 Claude agents ran 21 hours and surfaced an enzyme nobody had named
Key takeaways
- Around 950 Claude agents ran for roughly 21 hours and burned about 210 million tokens
- They narrowed more than 200,000 reverse transcriptases to 20 candidates and surfaced array-associated reverse transcriptases
- ART is a preprint finding with no wet lab confirmation, and nobody knows what it does yet
Around 950 Claude agents ran for roughly 21 hours and burned about 210 million tokens.
Anthropic launched a life sciences research group this week and published what its own model turned up on the way. Claude was pointed at a large genomic database with a single instruction: look for unusual reverse transcriptases.
What came out
The run surfaced more than 200,000 reverse transcriptases, generated about 3,500 candidate systems, and narrowed those to 20 worth a closer look. Out of that came array-associated reverse transcriptases, or ART, found in bacteriophages, the viruses that infect bacteria.
The enzyme itself was already known. What Claude flagged was the arrangement: a reverse transcriptase, a neighbouring partner gene, and a long array of evenly spaced DNA repeats sitting alongside them.
That repeat structure is why people paid attention. Evenly spaced repeat arrays are the feature that makes CRISPR programmable, because the spacing is what lets a system store and index sequences it has seen before. Finding the same architecture in an unrelated enzyme family is the kind of pattern that either means something significant or means nothing at all.
Why it matters
Nobody yet knows what ART actually does. The work is a preprint, not a peer-reviewed paper, and there is no wet lab confirmation. Treat the biology as an unverified lead.
The claim worth testing is the timeline. Anthropic says this class of database search normally takes expert scientists weeks to months. If 21 hours holds up as a repeatable figure rather than a favourable single run, the interesting result is not the enzyme, it is that hypothesis generation has become cheap enough to run speculatively. That is a different economics to the one behind Cognition's $2 billion Series E, where the bet is on agents finishing defined work rather than proposing new work.
It is also worth noting who is making the claim. Anthropic evaluated its own model, published its own preprint, and announced a commercial life sciences group in the same week. That does not make the result wrong. It does mean independent replication is the thing to wait for.
What to watch for next
The first signal is whether an outside lab reproduces the search and lands on the same 20 candidates. The second is whether anyone characterises what ART actually does inside a living cell. The third is cost, because 210 million tokens is affordable for Anthropic and not yet affordable for a university group, though the recent round of API price cuts is moving that line.
Until at least the first of those lands, the honest summary is that a very large search found something structurally odd, quickly.
