Thirty-two minutes, measured. Months to a year, estimated.
Start by separating the two numbers in the headline, because only one of them was measured. The 32 minutes is a wall-clock time: 1,946 agents, each handed one gene disruption from a whole-brain Perturb-seq atlas, each reading the cellular response, searching the literature and filing a report, all running in parallel. Parallelism is doing most of the work in that figure. Split any large reading job across enough parallel workers and the clock shrinks; the interesting question is what each worker produced.
The other number was not measured at all. OpenAI's announcement says the team estimated that comparable human analysis, done one perturbation at a time, would take months to a year. The preprint is blunter about its own figure: it calls the human side an illustrative estimate, built on assumptions of $22 an hour, 15 minutes to interpret one perturbation and 10 minutes per paper read, and states plainly that this is not measured human performance. No human team was timed, and no partial human run is reported for calibration. That does not make the estimate unreasonable, but an estimate offered by the team selling the result is not a benchmark, and it should not harden into one as the story travels.
There is also a scope point the headline compresses away. The agents analysed data that already existed. The expensive, slow part of this science, switching off 1,947 genes in living mice and measuring 7.7 million cells, was done by the Scripps team and published in March. The agents did the reading, not the experiment.
The comparison that was actually run
The study's controlled comparison is not agents against humans. It is many agents against one agent. The preprint held the model, the prompt, the reasoning setting, the tools and the data access fixed, then scaled the workload from one perturbation to 893, the subset of the atlas with detectable transcriptional responses. At that largest tested workload, the single agent's literature search thinned out until it cited 11 references. The distributed agents retained 711 times as many references, because each one only had to read deeply about a single gene. That is the gap the preprint's abstract summarises as hundreds of times more references. OpenAI's announcement phrases the distributed total differently, as more than 700 references cited, which does not match the preprint's own 711-fold figure; the two accounts quantify the same comparison differently. Neither version measures accuracy. A reference count shows how much literature a system touched, not how well it chose or used it.
The quality check points the same way, with a limit worth naming. Thirty doctoral-level experts each assessed six reports, 60 in all, without knowing which system wrote them, scoring novelty, evidential support, follow-up priority and confidence. In a separate matched comparison, 20 report pairs were each judged by three experts: the distributed reports were preferred in 45 of the 60 judgments, the single agent's in 10, with five ties. Preference between two machine-written reports tells you which read better to an expert. It does not tell you which was right, because for the new leads, nobody knows yet. The agents' own novelty scores tracked the experts' moderately, at a Spearman correlation of 0.61: agreement, not alignment.
The result that will travel least is in the same preprint. Asked to predict the effects of 272 held-out disruptions in one cell type, 100 that produced a transcriptional response and 172 that did not, the agents caught 59 of the responses and 142 of the non-responses, a balanced accuracy of 0.705. The preprint reports that this beat the baseline strategies it compared against, built from response distributions and physical-interaction neighbours. The class split still matters when reading the raw counts: a strategy of predicting no response every time would have been right 172 times out of 272, while sitting at chance on the balanced measure the paper uses. The preprint itself closes the loop on scope: evaluation on independent datasets and prospective experiments is what would establish how well the framework generalises.
Two leads, both still on paper
The output worth following is specific. Tsc1 and Tsc2 work as a pair to control mTOR, a central regulator of cell growth. In most neuron groups their disruptions looked similar. In the cerebellum they did not: Tsc2 disruption moved dozens of genes involved in neuronal communication, Tsc1 disruption barely registered. That lands on real human biology, because cerebellar problems in tuberous sclerosis are strongly tied to TSC2 mutations, and it gives the Jin lab a precise question to take back to animals.
The second lead is a hypothesis with a mechanism attached. Disrupting Dctn1, the gene mutated in Perry syndrome, turned up the activity of cholesterol-producing genes. The agents connected that to literature suggesting neurons compensate when cholesterol transport fails, and asked whether disturbed cholesterol distribution inside neurons feeds the degeneration. Experts in the study rated the idea novel and gave it a high follow-up priority. That is the correct status. The team presents these as hypotheses for further experimental investigation, and the sources reviewed do not report prospective laboratory validation of either newly proposed mechanism.
Read the byline before the abstract
None of this requires doubting the work to price it correctly. The preprint's competing-interest statement is easy to read: six of the authors are OpenAI employees and two are PerturbAI employees, and Grace X. Y. Zheng and Xin Jin are co-founders of PerturbAI. The model at the centre of the paper, GPT-Rosalind, is one OpenAI offers through a trusted-access programme rather than as a public product. A paper showing that product doing months of reading in half an hour is, in effect, also a demonstration of the product; that is our reading of its dual role, not the authors' description of it. And it has not yet been peer reviewed.
One number is absent from the accounts we verified: what the run cost. Neither OpenAI's announcement nor the preprint material checked for this article provides a reproducible total cost for the 1,946-agent run. The 32-minute wall-clock result reflects parallel execution, but without token consumption, tool-call counts, concurrency limits and compute or API pricing, readers cannot establish the run's total resource cost or compare its economics with a sequential workflow. Speed and cost efficiency are separate questions, and so far only one of them has a number attached.
What survives those discounts is still a genuine shift in where the bottleneck sits. Experiments now produce more interpretation work than the scientists who ran them can read, and this work is a team-reported demonstration that a swarm of agents can do the first pass and hand back a ranked list of questions. The ranking is the product. The answers still come from the lab, at lab speed.
Sources
- [1] OpenAI, "Using thousands of AI agents to uncover new leads in biological data", OpenAI, 21 October 2026 — the 32-minute run, the months-to-a-year estimate in the team's own wording, the single-agent versus distributed comparison at the largest workload (11 vs "more than 700" references), the 75% blinded-preference figure, and the Tsc1/Tsc2 and Dctn1 leadsRead source
- [2] Sun et al., "Distributed AI agents enable comprehensive discovery of biological insights at scale", bioRxiv preprint, posted 16 October 2026 (the DOI embeds the 15th) — read in full for this article: the 1,946 agents, GPT-Rosalind-5.5 and its December 2025 knowledge cutoff, the withheld source manuscript and the recovery of its testable findings, the 893-perturbation scaling comparison with the single agent's 11 references against 711 times as many retained by the distributed agents, the 3,397 synthesised findings, the blinded expert design (30 experts, 60 reports, 20 matched pairs, 45 of 60 preferences, novelty correlation 0.61), the held-out results (59 of 100 responses and 142 of 172 non-responses, balanced accuracy 0.705 against the preprint's baselines), the illustrative human-effort assumptions, and the competing-interest statement (six OpenAI employees, two PerturbAI employees, Zheng and Jin as PerturbAI co-founders)Read source
- [3] Shi et al., "Genome-scale functional mapping of the mammalian whole brain with in vivo Perturb-seq", bioRxiv preprint, 16 March 2026 — the underlying atlas: loss of 1,947 disease-associated genes profiled across more than 7.7 million cellsRead source
- [4] The Decoder, "OpenAI agents analyze mouse brain data in 32 minutes, a job that would take humans months", 22 October 2026 — independent same-day summary confirming the announcement's scope and the GitHub reasoning atlasRead source