OpenAI agents on GPT-Rosalind analyze mouse brain CRISPR atlas in 32 minutes, versus months by hand

OpenAI, PerturbAI and Scripps Research ran thousands of AI agents on a whole-brain CRISPR atlas of mouse cells and completed the analysis in 32 minutes. The team estimates the same work would take human researchers anywhere from several months to a year. The figures come from OpenAI; the accompanying paper is on bioRxiv and has not been described as peer reviewed.

What was done

According to OpenAI, the agents ran on GPT-Rosalind and worked through a dataset covering millions of mouse cells. The dataset is a whole-brain CRISPR atlas, which maps how gene changes affect different cell types in the mouse brain.

The workload was split by gene perturbation:

• Each agent took on a single gene perturbation.
• Each agent analyzed how the cells responded to that perturbation.
• Each agent searched the scientific literature for possible explanations.

The full run took 32 minutes. The manual-analysis estimate of several months to a year is the research team's own and was not independently measured.

Findings and evaluation

OpenAI says the agents surfaced, among other results:

• Unexpected differences between two disease-related genes.
• A possible link between cholesterol regulation and neurodegeneration.

In blind comparisons, human experts preferred reports from the distributed agents over reports from a single agent 75 percent of the time. The comparison is against a single-agent baseline, not against human-written analyses.

The output, called the "Reasoning Atlas," is available on GitHub. OpenAI says the leads should help researchers decide which mechanisms to test in the lab first.

What is not disclosed

The source material does not give:

• The exact number of agents (described only as "thousands").
• GPT-Rosalind's context window, parameter count, training cutoff or pricing. These are not yet disclosed.
• Compute cost for the 32-minute run.
• Any wet-lab validation of the reported findings.

What this means

The notable result is the architecture, not a new model. Splitting a large biological dataset into one-agent-per-perturbation units turns a serial literature-and-analysis task into a parallel one, so wall-clock time depends mostly on available compute rather than analyst hours. The 75 percent preference over a single agent is the most useful data point here, because it suggests that distributing the work improves report quality as well as speed.

The caveats matter. The months-to-a-year baseline is an estimate. The preference test does not compare agents with human analysts. The findings are hypotheses that still need lab testing, which OpenAI itself frames as prioritization rather than confirmation. Until the bioRxiv paper is reviewed and the Reasoning Atlas is checked by outside groups, the 32-minute figure shows throughput, not validated discovery.

For builders, the pattern is reusable: partition by a natural unit of the problem, give each agent retrieval over the literature, and aggregate the outputs. How much of the result depends on GPT-Rosalind specifically is unclear until OpenAI publishes more detail on the model.