- An OpenAI speaker said 10,000 agents, 2.7 million messages and about 130 billion output tokens went into the full effort on one maths problem.
- The model was internal and not named, and the excerpts show no outside check, so readers rely on OpenAI's own account.
At OpenAI's DevDay 2026 session "What Comes After Loops?", a speaker on the show said the company had found a solution to the Navier-Stokes problem. It is a long-standing Millennium Prize problem in mathematics. The speaker said 10,000 agents worked on it together. The figures show how much compute one claimed result can now involve. They also show how little outsiders can check when the model is internal and unnamed.
What was said
The speaker said OpenAI used an internal model and Codex and found a solution with 10,000 agents working together. The excerpts do not show anyone outside confirming it. Across the full effort, the speaker said, "there were 2.7 million messages and about 130 billion output tokens." Tokens are the units of text that AI services bill for. This is a total for the whole effort, not a stated minimum the result required.
Teams started with different versions of the problem and were prompted to try different approaches. They could share findings and ask other teams questions. The aim was to stop duplicate work, because having all 10,000 agents on one thing would be very expensive.
When the teams found what the speaker called the unforced Euler result, all teams were turned toward solving the problem. The speaker also said the internal model had resolved more than 100 long-standing maths problems.
Why it matters
Our reading: the compute is part of the claim, and the checking is the open part. The speaker gave no cost, so it is unclear who could afford to repeat this work. The model was described only as internal, so outsiders may have to lean on the vendor's own account.
If you sign off on AI-made work, ask what it cost to produce. Ask also what an independent party could actually rerun.
The other side
The speaker pointed the audience to a QR code with more detail on the Navier-Stokes work, so more material may exist than the talk itself. The excerpts do not say whether outside mathematicians have checked the result.
Another speaker, Eric, warned in the same session that "Every agent Codex brings in can dramatically increase token use without necessarily making the work faster or better." His warning concerned everyday tasks, not the maths effort. But it is a reason not to read the total as the cost of the answer. That link is our inference. A conference talk is not a peer review.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- OpenAI: What Comes After Loops?: Deep Dive Into Agentic Architecture | DevDay 2026 (2026-11-11)
