OpenAI wrote research pace out of its new math advisory group's remit
The company says an internal model cleared more than 100 open problems after a month of training, and the advisory group it is backing at the Institute for Advanced Study has been handed the release schedule to work on.
The Engineer · Build desk
What happened
- OpenAI says a new internal model has solved more than 100 long-standing problems across most areas of mathematics, on top of its Navier-Stokes Millennium Problem solution.
- Most of those 100-plus results have not been published, so mathematicians outside the company cannot yet judge their quality or novelty.
- The Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, says its current task is advising OpenAI on how to coordinate the release of that large batch of results.
- Fields Medalist Timothy Gowers is a founding member of the group and did not sign the open letter from the 25 Fields Medalists criticising open-problem benchmarks.
Why it matters
- What it costs cost The checking work falls on academics OpenAI does not pay, and it grows with every batch the company decides to release.
- Who has to decide decision Anyone pricing the 100-plus figure has to treat it as an internal count until the proofs are out and reviewers have decided what each one adds.
- What it sets up precedent Expert oversight scoped to disclosure, with research speed carved out in writing, is now a documented template other labs can copy when specialists object.
- What doesn't add up contradiction OpenAI's chief scientist recently said the team deliberately chose not to optimize for math; the announcement now leads with math results.
Next story · Science · 2d ago
Thomas Bloom stops marking Erdős problems open or solved as AI-written proofs multiplyConfirmed2 publishers
"Importantly, the group will not be responsible for advising us on how to pace our internal progress on mathematics," OpenAI writes [7]. Members are unpaid, they pick who joins, they can offer advice nobody asked for, they can speak publicly about the company's influence on mathematics, and they can publish their recommendations [6]. The Neuron reports that the group has no decision-making authority over OpenAI [8].
A publication right without a decision right means the group can put a disagreement on the record and OpenAI can proceed anyway. The Neuron notes that a slower release schedule gives mathematicians more time per result and does not reduce the rate at which new work enters OpenAI's internal queue [13].
The one run OpenAI attached numbers to is Navier-Stokes. The company says that effort used roughly 10,000 concurrent agents with access to cached internet material and code execution, produced the proposed solution in about 88 hours, and then took roughly 17 hours of Lean formalization [10][4]. That is 105 hours, about 4.4 days [21]. Training began on August 28 and the announcement came on September 21, a window of 24 days [2][22]. Five runs of that length fit end to end in 24 days. Reaching 100 results inside the same window would need roughly 18 of them at once, and at 10,000 agents each that is about 180,000 concurrent agents [23]. That assumes the other results cost what Navier-Stokes cost. The company has not released most of them and did not say how long those runs took [3].
OpenAI published the Navier-Stokes proof in prose and in Lean, which checks reasoning against defined assumptions [11]. Reviewers then do the parts Lean cannot: confirm the proof answers the intended problem, check whether someone already established part of the result, and identify which ideas are actually new [12]. The published solution has already sparked heated debate in parts of the scientific community, according to the-decoder [19].
The September 11 declaration "A Severe Misalignment of AI in Mathematics" argues against using famous open problems as benchmarks for AI systems, on the grounds that progress also depends on understanding why a result works and whether its ideas can support further research [9]. Timothy Gowers agrees with much of the letter and thinks mathematics is in trouble. Where he splits from it is on what mathematics is for: the letter treats conceptual understanding as the goal and problem-solving as the means to it [16]. He set that out on his blog [14].
According to the-decoder, OpenAI acknowledged that it took on the Millennium Prize Problem only after hearing rumors that another team, partly made up of Anthropic researchers, had already cracked it [18]. OpenAI calls the collaboration a "first step," with "difficult questions ahead about how AI can support mathematical understanding and how the benefits of these capabilities can reach the wider community" [20].
What to watch
- Whether OpenAI publishes the remaining results with Lean artifacts, and what release schedule the advisory group recommends.
- Whether independent mathematicians confirm the Hodge conjecture result reported as among the solved problems.
- Whether the group uses its publication right to disagree in public with an OpenAI release decision.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+55
- Incentives70
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
OpenAI says a new internal model has solved more than 100 long-standing problems across most areas of mathematics, on top of the Navier-Stokes Millennium Problem.
- [2]
The model began training on August 28, based on OpenAI's September 21 announcement.
- [3]
OpenAI has not released most of the 100-plus results publicly, so outside researchers cannot yet judge their quality or novelty.
- [4]
The system generated the proposed Navier-Stokes solution in about 88 hours, followed by roughly 17 hours of Lean formalization.
- [5]
The Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, says its current task is to advise OpenAI on how to coordinate the release of a large number of mathematical results produced by the company's internal model.
- [6]
OpenAI says the group will operate independently: members can offer advice without being asked, speak publicly about the company's influence on mathematics, and publish their recommendations; OpenAI does not pay them and they can decide who joins.
- [7]
"Importantly, the group will not be responsible for advising us on how to pace our internal progress on mathematics," the company writes.
- [8]
The group does not have decision-making authority over OpenAI, and its role does not include controlling how quickly the company continues its internal mathematical research.
- [9]
In the September 11 declaration "A Severe Misalignment of AI in Mathematics", mathematicians including Fields Medal winners criticized using famous open problems as benchmarks for AI systems, arguing that progress also depends on understanding why a result works and whether its ideas can support further research.
- [10]
OpenAI says the successful research effort used roughly 10,000 concurrent agents with access to cached internet material and code execution.
- [11]
OpenAI published both a written proof of the Navier-Stokes result and a version in Lean, software that can mechanically check mathematical reasoning against defined assumptions.
- [12]
When an AI produces a proposed proof, mathematicians must confirm that the proof answers the intended problem, determine whether someone has already established part of the result, and understand which ideas are actually new.
- [13]
A controlled release schedule may give mathematicians more time to review each result, but it does not reduce the rate at which new work enters OpenAI's internal queue.
- [14]
Timothy Gowers did not sign the open letter from the 25 Fields Medalists and laid out his reasoning on his blog.
- [15]
One of the group's founding members is Timothy Gowers, the prominent mathematician and Fields Medalist.
- [16]
Gowers agrees with much of the letter and thinks mathematics is in trouble; where he parts ways is on what mathematics is for, as the letter treats conceptual understanding as the main goal with problem-solving just a means to get there.
- [17]
Chief Scientist Jakub Pachocki recently said the team had deliberately chosen not to optimize for math and was focused on recursive self-improvement instead.
- [18]
OpenAI admitted it only took on the Millennium Prize Problem after hearing rumors that another team, partly made up of Anthropic researchers, had already cracked it.
- [19]
The published Navier-Stokes solution has already sparked heated debate in parts of the scientific community.
- [20]
OpenAI calls the collaboration a "first step," with plenty of "difficult questions ahead about how AI can support mathematical understanding and how the benefits of these capabilities can reach the wider community."
- [21]
The Navier-Stokes run took about 105 hours in total, or roughly 4.4 days.
- [22]
August 28 to September 21 is a window of 24 days.
- [23]
If every result cost what the Navier-Stokes run cost, about five runs fit end to end in the 24-day window, so producing more than 100 results would require roughly 18 runs in parallel, or about 180,000 concurrent agents.
- [24]
Among the solved problems is reportedly a second Millennium Prize Problem, the Hodge conjecture.
Sources
2 independent publishers whose own reporting we read for this story.
- the-decoder.comOpenAI says its internal model solved over 100 long-standing math problems after just a month of training
2 articles · September 22, 2026
- theneuron.aiOpenAI Says Its AI Solved 100+ Open Math Problems. Can Researchers Keep Up?
2 articles · September 21, 2026
Topics and entities
Entities
- OpenAI
- Advisory Group on Mathematics and Artificial Intelligence
- Institute for Advanced Study
- Hodge conjecture
- Lean
- Jakub Pachocki
- Anthropic
- Navier-Stokes existence and smoothness problem
- A Severe Misalignment of AI in Mathematics