OpenAI agents mine mouse brain CRISPR atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research said on October 21, 2026, that thousands of GPT-Rosalind agents running in parallel interpreted an existing whole-brain mouse CRISPR atlas covering more than 7.7 million cells in 32 minutes. OpenAI's blog post calls the method "distributed biological reasoning." The output is a public "reasoning atlas" of candidate gene functions and disease leads, described in a new bioRxiv preprint.

Beyond this one dataset, the stakes are clear. Genomics labs now generate perturbation data faster than scientists can read it, so interpretation has become the bottleneck. OpenAI also needs proof that its paid biology model does real work. GPT-Rosalind left research preview on September 11 and gained published pricing on October 5.

Inside the 32-minute run

The source data is not new. It comes from a Perturb-seq atlas posted in March by Xin Jin's group, with authors from Scripps Research, PerturbAI, HHMI, NVIDIA, the Allen Institute and Salk. That study disrupted 1,947 disease-associated genes in living mouse brains. PerturbAI has posted the data on Hugging Face under a CC-BY-4.0 license.

In the new work, each agent took one gene disruption, examined how cells responded, searched the literature and proposed explanations. The reports were merged into an atlas linking genes, cellular effects, publications and possible mechanisms. OpenAI says only "thousands" of agents took part; R&D World puts the figure at 1,946. The code is on GitHub under an MIT license, with a hosted site and a Codex MCP server.

The 32-minute figure needs context. It covers analysis only, not producing and sequencing the cells or running follow-up experiments, a point the Italian outlet 01net also makes.

The 'months to a year' comparison

OpenAI says doing the same analysis by hand, one perturbation at a time, would take "months to a year." That claim is softer than it looks. According to R&D World, the estimate is illustrative: it assumes 15 minutes per gene disruption and 10 minutes per paper. No team of human analysts was tested head-to-head against the system.

The stronger comparison is internal. Holding the model and compute fixed, OpenAI says a single agent cited only 11 references at the largest workload, while the distributed agents cited more than 700. Blinded experts preferred the distributed reports 75% of the time. R&D World reports that 30 doctoral-level biologists did the review, and that the agents matched or refined 48 of the 55 findings drawn from their data without seeing the original atlas paper.

No benchmark against rival science-agent systems was offered. The model itself has a track record: at its April launch, OpenAI said GPT-Rosalind led BixBench among models with published scores and beat GPT-5.4 on 6 of 11 LABBench2 tasks. Amgen, Moderna, the Allen Institute and Thermo Fisher Scientific were named as customers.

Tsc2, Dctn1 and a cholesterol hunch

OpenAI highlights two leads. In the cerebellum, disrupting Tsc2 changed dozens of genes tied to neuronal communication, while disrupting Tsc1 changed little. OpenAI connects this to tuberous sclerosis, a genetic disorder in which cerebellar abnormalities are strongly linked to TSC2 mutations.

The second lead involves Perry syndrome, a neurodegenerative disease with Parkinson-like symptoms caused by DCTN1 mutations. In the atlas, disrupting Dctn1 increased the expression of cholesterol-synthesis genes. That led to an untested hypothesis that DCTN1 loss changes how cholesterol is distributed inside neurons. Three blinded reviewers gave it an average follow-up priority of 4.3 out of 5, per R&D World.

"Sometimes the clues are already there, scattered across different fields—we just haven't connected them," said Tomasz Nowakowski of UCSF in OpenAI's post. R&D World notes he is not an author of the preprint. Sarah Teichmann, founding co-chair of the Human Cell Atlas, said AI "could help more researchers explore these resources." Both comments appear in OpenAI's own announcement, and no independent critic has yet commented on the record.

Preprint authors sound more cautious. As quoted by R&D World, they write that "reading the literature, weighing explanations, and choosing decisive experiments would remain central to scientific training and judgment."

Timing matters too. OpenAI listed a Rosalind Academic Plan, dated October 15, alongside the announcement, which suggests a push for university users, though the company has given no reason for the timing.

Watch the lab next. The key question is whether any group tests cholesterol distribution in Dctn1-deficient neurons, and whether independent teams use the open code to reproduce the 48-of-55 result. Until then, the reasoning atlas is a list of well-organized hypotheses, not discoveries.