OpenAI announced on October 21 that 1,946 of its GPT-Rosalind agents, running in parallel, analyzed existing data from millions of mouse brain cells in 32 minutes — work the team estimates would take a human research group months to a year to complete through manual analysis and literature review.
The demonstration, conducted with drug-discovery company PerturbAI and Scripps Research, is described in a bioRxiv preprint and applies what the team calls "distributed biological reasoning" to a whole-brain CRISPR atlas — a dataset cataloging how disrupting individual genes changes gene expression across brain cells. For more context on this story, see our ongoing artificial intelligence updates.
How Distributed Biological Reasoning Works
CRISPR perturbation experiments let researchers switch off genes and measure the resulting cellular changes, but they can generate more observations than any single lab can investigate in depth. That bottleneck is precisely why researchers share cell atlases publicly. PerturbAI, the drug-discovery company that co-developed the approach, and Scripps Research supplied the scientific framing and domain review for the collaboration.
The software described in the preprint assigns each gene disruption to an independent AI agent. Each agent examines the cellular changes associated with its assigned gene, searches the scientific literature, and proposes explanations for what it finds. A coordination layer then combines the agents' reports, identifies connections across the full dataset, and ranks the most promising ideas for human follow-up.
"We build cell atlases so that other scientists can make discoveries we haven't imagined yet," said Sarah Teichmann, chair in stem cell medicine at the University of Cambridge and founding co-chair of the Human Cell Atlas, in OpenAI's announcement. "AI could help more researchers explore these resources and turn unexpected observations into new questions to test."
The Numbers: 711 Times the References, 48 of 55 Findings Recovered
The scale difference between one agent and nearly 2,000 is stark. In a comparison using the same model, prompt, tools, and data access, single-agent runs averaged about 11 literature references at the largest workload, while the distributed system retained 7,586 references — approximately 711 times as many, according to the project's public data.
The agents also held up well against the original science. Working without access to the atlas study's manuscript, they matched or refined 48 of the original study's 55 findings that could be drawn from the data the agents received.
To assess quality, the team recruited 30 doctoral-level biologists for a blinded review. Their assessments broadly supported the agents' rankings of the most promising research leads.
The DCTN1 Cholesterol Lead
One finding illustrates the kind of hypothesis the system is meant to surface. Disrupting DCTN1 — a gene linked to neurodegeneration — increased the activity of genes involved in cholesterol production. The single-agent baseline proposed that problems transporting cholesterol within neurons might prompt the cells to compensate by producing more.
Three blinded reviewers gave that lead an average follow-up priority score of 4.3 out of 5. Tomasz Nowakowski, an associate professor of neurological surgery at UC San Francisco who was not an author of the preprint, said in OpenAI's announcement that the finding shows why linking observations across disciplines matters: "Sometimes the clues are already there, scattered across different fields — we just haven't connected them."
Promise and Caveats
The announcement arrives amid rapid momentum in OpenAI's life sciences push. GPT-Rosalind, introduced earlier this year, is a model specialized for life sciences research that combines biological reasoning with scientific tool use, with documented applications spanning genomics, protein analysis, target research, medicinal chemistry, and experiment planning. The company has also rolled out a Rosalind Workbench for scientific workflows and a Life Sciences Research plugin in Codex, and in June it introduced LifeSciBench, an evaluation the company says measures research assistance rather than proving that a system can independently validate discoveries or replace experimental verification — a distinction that carries over to the current announcement.
Important caveats remain. The 32-minutes-versus-a-year comparison is an illustrative estimate built from assumptions — 15 minutes per gene disruption and 10 minutes per paper read — rather than a measured equivalent. The work is a proof of concept: the AI surfaced hypotheses for humans to test, not validated discoveries, and none of the leads have been experimentally confirmed. The speed comparison with human analysts likewise rests on the team's illustrative time assumptions rather than a head-to-head trial, and the expert reviewers graded the agents' reports rather than racing them against a human research team. Independent replication of the benchmark-style comparisons has not yet occurred.
Still, the demonstration sketches a plausible future workflow for biology: public datasets endlessly re-mined by swarms of specialized agents, with human scientists spending their time testing and refining what the machines rank highest. Whether those rankings consistently hold up in the lab will determine whether distributed AI reasoning becomes standard equipment in the life sciences.
Source note: Details in this article are drawn from OpenAI's October 21, 2026 announcement and coverage by R&D World, including the underlying bioRxiv preprint (Sun et al., 2026).---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
