OpenAI published a research note on October 21, 2026 describing a run in which thousands of AI agents, powered by a model it calls GPT-Rosalind, worked through existing data from millions of mouse brain cells in 32 minutes. The work was done with PerturbAI and Scripps Research and is called distributed biological reasoning.
What was actually released
A method, a write-up and a public atlas. There is no model to buy here and no plan: this is a research publication, and the only artifact a reader can open today is the reasoning atlas the agents produced, which the team says is public.
Each agent took one gene perturbation from a whole-brain CRISPR atlas, the dataset that records how disrupting a gene changes different cell types across the mouse brain. The agent looked at the cellular response, searched the scientific literature, and proposed an explanation. The reports were then merged into one structure that links a gene, its cellular effects, the papers that speak to it, and a candidate mechanism. Thousands of those reports, run in parallel, are what produced the 32 minutes.
Two findings carry the note. The first is a difference between Tsc1 and Tsc2, two genes that both control mTOR. Disrupting either had similar effects in several groups of neurons, but in the cerebellum Tsc2 disruption changed the activity of dozens of genes involved in neuronal communication while Tsc1 disruption produced little change. OpenAI ties that to tuberous sclerosis, where cerebellum abnormalities are linked to TSC2 mutations, and to the question of which existing medicines might correct which changes.
The second is DCTN1, the gene behind Perry syndrome. Disrupting it increased the expression of genes involved in making cholesterol, which the agents connected to published work on cholesterol transport. That produced a question rather than an answer: whether DCTN1 disruption changes how cholesterol is distributed inside neurons, and whether restoring that distribution would protect them.
“We build cell atlases so that other scientists can make discoveries we haven’t imagined yet. AI could help more researchers explore these resources and turn unexpected observations into new questions to test.”
Sarah Teichmann, chair in stem cell medicine at the University of Cambridge and founding co-chair of the Human Cell Atlas, quoted in the OpenAI note.
What the claim rests on
The most useful part of the note is not the biology. It is the one comparison the team says it ran with the model and the resources held fixed: a single agent given the entire dataset against the distributed agents at the largest workload. As the workload grew, the single agent searched less of the literature and cited 11 references, while the distributed agents cited more than 700.
| What the note reports | Figure |
|---|---|
| Time to work through the mouse brain atlas | 32 minutes |
| References cited, distributed agents, largest workload | more than 700 |
| References cited, single agent, same model and resources | 11 |
| Blinded comparisons where experts preferred the distributed reports | 75% |
| The team’s estimate for the same review done by hand | months to a year |
Two of those rows are measurements and one is an estimate. The reference counts come from a run the team describes, and the 75% comes from blinded comparisons where experts ranked reports without knowing which setup produced them. The comparison with human work is neither: the months-to-a-year figure is the research team’s own estimate, not a timed benchmark. It is a fair estimate, and the note does not present it as a measurement.
What is missing is as informative. OpenAI does not publish what the run cost: no token count, no model spend, no price for GPT-Rosalind, no compute figure for thousands of parallel agents. It does not say how many agents “thousands” means, how many perturbations were covered, how many experts sat on the blinded panel, or how the single-agent baseline was scaled. It reports no false-lead rate, and does not say whether either highlighted mechanism has been tested at the bench. The paper behind the work is a preprint, not a peer-reviewed article.
None of this makes the result wrong. It makes it a demonstration that the method produces plausible, expert-endorsed leads at speed, which is a different claim from the leads being correct.
What can be checked today
The reasoning atlas is public, so researchers can query the gene-to-effect links themselves. Everything else is a description of a run outsiders cannot audit: no logs, no costs, no per-lead scoring.
On this page
What it changes for teams buying agent tools
Vendors sell agent products on a demo and a price. This note suggests the question between the two: what happens to the same model when you stop giving it more context and start giving it more workers.
The 11-against-700 comparison is the version of that question that can actually be answered, because the model and resources are fixed and only the shape of the work changes. The failure of the single agent is not that it invents things. The note says the failure is that it searches less, so it sees less of what the data already contains. That is the mode buyers rarely test for: a thin report reads as confidently as a broad one.
Three things follow for anyone running agents in production. Fan-out is a retrieval strategy before it is a compute strategy, and the number worth instrumenting is coverage, not speed. A benchmark that reports only the distributed result is a demo. And the deliverable matters: they shipped a browsable atlas others can query, not a screenshot of an answer.
The Neural Minute take
We have spent the autumn writing about agents that overstate what they did, and the InnovationEval result is the one we keep returning to: agents that burn real compute and then report success they cannot support (our coverage). This note is unusually careful. It puts the estimate in the same paragraph as the measurement and says which is which, and it publishes the comparison in which its own method could have lost.
So the number to keep is not 32 minutes. It is 11 references. A lab published the case where a bigger context window loses to more workers on the same model, with the count that shows it, against the orchestration pitches we covered this month (Claude Managed Agents). Unlike most such claims, this one is reproducible on your own corpus: split the job, hold the model fixed, count what each side reads.
The reservation is the bill. Thousands of agents in parallel for 32 minutes is cheap only if you know what it cost, and OpenAI does not say. If the run cost more than the human work it replaces, the method is a research instrument rather than a production tool, and 32 minutes is a capability claim rather than a price claim. Ask for cost per result before you buy the fan-out story, here and everywhere else agent swarms are sold this quarter.
What the note does not publish
Cost per run, token or compute spend, the number of agents behind “thousands”, the size of the blinded expert panel, a false-lead rate, and any bench validation of the two leads it highlights. The paper behind the work is a preprint.
Sources
- Using thousands of AI agents to uncover new leads in biological data, openai.com, published October 21, 2026: https://openai.com/index/scaling-ai-research-for-complex-biological-data
- All figures above come from that page: 32 minutes, more than 700 references against 11, 75% of blinded comparisons, and the team’s estimate of months to a year. The two gene findings and the expert comments are as written there. No other publisher was used.
- Related Neural Minute coverage: https://www.neuralminute.com/news/ai-agents-overstate-their-results-and-remain-far-from/ and https://www.neuralminute.com/news/anthropic-s-claude-can-now-orchestrate-up-to-1-000-ai/