Claude agents narrowed more than 200,000 reverse transcriptases to 20 candidates and led researchers to a previously unrecognised genomic system. Unlike several earlier AI-for-science claims, the lead went to the lab and held up.
Anthropic researchers used Claude agents to identify an array-associated reverse transcriptase, or ART, in a bacteriophage. The enzyme family was already known. The new result was the surrounding system: a repeating DNA array, short RNAs expressed from its repeats and an adjoining partner gene whose function remains unknown. Laboratory work confirmed the RNA expression; the system’s biological role is still open, and the work is a preprint. Anthropic’s research account describes the search and experiments.
The team began with more than 200,000 reverse transcriptases, identified about 3,500 candidate systems and had Claude analyse 20 in detail. Each report combined sequence evidence with genomic context and proposed a function. Researchers took a standout system to the lab and confirmed short-RNA expression. The close-review set was about 0.01% of the starting count: that is how far the search shrank the pool before anyone touched a pipette.
The workflow split the labour cleanly. Agents surveyed a protein family, read relevant papers, reproduced known findings against public data and then looked for family members near genomic features that did not fit established systems. Human scientists set the question, reviewed the reports, chose what to test and interpreted the experiments. The agents sustained a wide first pass across candidates and literature; the lab decided which leads held up.
The computing bill was small. Anthropic reports that roughly 950 agents ran for 21 hours and used 210 million tokens, about 19,950 agent-hours if every instance ran continuously. Sonnet 5’s listed API prices put the token cost between $420, if all tokens were billed as input, and $2,100, if all were output. U.S. biochemists and biophysicists had a median annual wage of $127,410 in May 2025; applying the Bureau of Labor Statistics’ private-industry compensation-to-wage ratio gives a loaded cost of about $3,500 for a median researcher-week. The token bill was 12% to 60% of that. Even at the top of the range, the whole screen cost less than one researcher-week, against the weeks to months Anthropic says conventional analysis of this kind can take. Anthropic’s rate card, BLS wage data and BLS compensation data provide the inputs.
What sets the result apart from other AI-for-science claims is that the lead survived contact with the bench. The field has produced three kinds of claim, and they are not equal.
The weakest are claims of scale without validation. DeepMind’s GNoME reported 2.2 million new crystal structures in 2023, hundreds of thousands of them predicted to be stable. A review in Chemistry of Materials the following year found “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.” The model generated candidates; the value of the list was never established in a lab. The critique sets out the objections. Sakana’s AI Scientist sits nearby: one of its machine-written papers passed peer review at a 2025 conference workshop, but it involved no experiment, humans chose which papers to submit, and the paper was withdrawn after review. A workshop acceptance tests writing, not discovery. Sakana’s account describes the submission.
The second kind reproduces something humans already knew. Google’s AI co-scientist, announced in February 2025, proposed a mechanism for how certain genetic elements spread between bacteria that matched unpublished experiments at Imperial College. That was an impressive check on the system’s reasoning, but it rediscovered a result rather than producing a new lead. Its other proposals, drug candidates for acute myeloid leukaemia and liver-fibrosis targets, were tested only in cell lines and organoids. Google’s account describes the work.
The third kind proposes something new and tests it. FutureHouse’s Robin system suggested in May 2025 that ripasudil, a glaucoma drug, could treat dry age-related macular degeneration; laboratory work confirmed the proposed cellular effect within about ten weeks. Anthropic’s result sits in this group. It is further from any application than Robin’s, since it describes a biological system rather than a drug, but it is a new finding confirmed by experiment rather than a candidate list or a rediscovery. FutureHouse’s announcement describes Robin.
That ranking gives labs a practical test for the next claim: did the lead go to the bench, was it new, and did it hold up? By that standard the enzyme result is small but real.
A human-readable report is what makes the screen useful. It links sequence features, neighbouring genes and prior findings into a hypothesis a scientist can inspect and reject. Researchers can compare more candidates before committing reagents and equipment time, then test whether a system is expressed, identify its substrate and measure its function. The gain sits before the experiment, where cheap computation directs scarce lab capacity toward a stronger lead.
The biology is still early. The RNA’s target, the partner gene’s function and the system’s effect on the phage remain to be tested. That places the result at basic discovery, well before evidence of a therapeutic or commercial use.
The cost arithmetic moves the constraint. When a screen of 200,000 sequences costs less than a week of one scientist’s time, candidate generation stops being scarce and assay capacity becomes the limit. A lab that already has a queue of unrun experiments gains little from a longer list. The labs that benefit will be the ones that can screen, reproduce and reject candidates quickly, and the right measure of these systems is validated discoveries per lab-week, not candidates per dollar.