Math advisory group announcements do not usually arrive attached to a capability claim, but on 21 September 2026 OpenAI published both at once. The company said an internal model had resolved more than 100 long-standing open problems across most areas of mathematics, and that it was working with mathematicians who had established an independent advisory body hosted at the Institute for Advanced Study in Princeton.
The two halves of that announcement point in opposite directions. One is a statement of speed. The other is a structure for slowing down, or at least for being seen to consult. Ten days earlier, 25 Fields Medallists had published a declaration arguing that the goals of the companies building artificial intelligence and the goals of mathematics are “severely misaligned.”
This article sets out what the math advisory group is, what OpenAI published about the model behind the claim, what the declaration actually says, and one detail about the overlap between the two name lists that the coverage so far has got wrong. The primary documents are all public, they are all short, and they do not quite agree with each other. Where they disagree, the disagreement is the story.
Table of contents
- What the Math Advisory Group Actually Is
- The Overlap Nobody Reported Correctly
- The Numbers OpenAI Published Before the Math Advisory Group Existed
- The Declaration That Preceded the Math Advisory Group
- What the Math Advisory Group Is Not Allowed to Do
- The Twenty-Four Day Timeline
- What the Math Advisory Group Means Outside Mathematics
- Frequently Asked Questions About the Math Advisory Group
- References
What the Math Advisory Group Actually Is
The announcement is short, and most of the substance of the math advisory group sits in what it is permitted to do rather than in what it is for.
The stated purpose
OpenAI’s framing is that the math advisory group “will serve as a bridge to the mathematical community and broader public, giving mathematicians a voice in how we move forward.” Its remit is the review and communication of emerging results: assessing significance, advising on how to coordinate dissemination, and advising on academic and professional standards of mathematical research.
The nine initial members
The math advisory group is hosted at the Institute for Advanced Study and has nine founding members drawn from Europe and the United States. Three of them hold Fields Medals — Timothy Gowers (1998), Edward Witten (1990) and Martin Hairer (2014). Camillo De Lellis and Witten are both at the Institute for Advanced Study itself.
The independence terms
OpenAI says the body will operate independently, may offer advice that was not requested, may comment publicly on the company’s impact on mathematics, and may make its advice public. Members are not paid by OpenAI, and the group controls its own membership. “Its value depends on its members being able to exercise their own judgement and challenge ours.”
What the Institute for Advanced Study said
The IAS added a qualifier about the math advisory group that OpenAI’s own post does not contain: “Although we will give advice, we do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company.” That sentence is the honest description of the arrangement.
| Member | Institution | Fields Medal | Name on the declaration |
|---|---|---|---|
| Martin Hairer | EPFL, Imperial College London | 2014 | Yes, initial signatory |
| Camillo De Lellis | IAS, GSSI | No | Yes, endorser |
| Ravi Vakil | Stanford | No | Yes, endorser |
| Timothy Gowers | College de France, Cambridge | 1998 | No |
| Edward Witten | IAS | 1990 | No |
| Francois Charles | ENS-PSL | No | No |
| Nikhil Srivastava | Berkeley, Simons Institute | No | No |
| Ulrike Tillmann | Oxford, INI | No | No |
| Melanie Matchett Wood | Harvard | No | No |
The Overlap Nobody Reported Correctly
The last column of that table is the part worth dwelling on, because it is checkable against two public lists and it has been reported wrongly.
What the coverage said
TechCrunch reported that only one of the nine — Camillo De Lellis — had also signed the mathematicians’ letter. That is understandable, because the declaration presents two separate lists: an initial signatory list of Fields Medallists on the front page, and a searchable endorser table behind it. Checking a math advisory group member against only the first one produces exactly that answer.
What the two lists actually contain
Searching both on 22 September 2026 gives a different answer. Martin Hairer appears on the Fields Medallist list as an initial signatory. De Lellis and Ravi Vakil both appear in the endorser table, at the Institute for Advanced Study and Stanford respectively. Gowers, Witten, Tillmann, Srivastava, Charles and Matchett Wood appear on neither.
Three of nine, not one of nine
So a third of OpenAI’s new math advisory group had already put their names to a document describing the company’s benchmark-chasing as detrimental to mathematics. That is a materially different picture from one outlier. It makes the body harder to dismiss as a friendly panel, and it makes its inability to comment on pacing more pointed.
Why the count moved
There is a second discrepancy in the same documents. Every report says 25 Fields Medallists signed. The archived version of the declaration, deposited under DOI 10.5281/zenodo.22737750 with a publication date of 11 September, lists exactly 25 authors. The live page at mathandai.org now lists 27. The two names added after the freeze are Vladimir Drinfeld (Fields Medal 1990) and David Mumford (Fields Medal 1974).
The Numbers OpenAI Published Before the Math Advisory Group Existed
The capability claim did not arrive with the math advisory group on 21 September. It arrived on 8 September, with an unusual amount of operational detail attached.
The Navier-Stokes result
OpenAI published a proof that the three-dimensional Navier-Stokes equations can develop a singularity in finite time, resolving statements “C” and “D” in the official Millennium Prize formulation. The solution describes a vortex that spirals inward and elongates while its energy stays finite. The company said it would not claim the $1 million prize.
The compute behind it
The figures OpenAI disclosed are the most concrete thing in this entire story. This is not a large language model answering a prompt; it is a fleet of agents coordinating over days. A new internal model began training on 28 August. The effort launched on 1 September after the company heard rumours that two Millennium Prize problems had been resolved. The winning group involved on the order of 10,000 concurrent agents, and reached its resolution on 5 September, about 88 hours later. Lean formalisation and verification took a further 17 hours on GPT-6 Astra.
What the message counts divide into
Across all attempted problems the agents sent 4.9 million messages and used about 300 billion output tokens. The Navier-Stokes problem alone accounted for 2.7 million messages and roughly 130 billion output tokens. Dividing those pairs gives the share of the whole effort that one problem consumed: 55.1% of the messages and 43.3% of the tokens.
The smaller result that came first
Before Navier-Stokes, the same system resolved the regularity problem for the Euler equations in the unforced case. That took nearly 100 agents working for about 50 hours. Multiplying agents by hours gives roughly 5,000 agent-hours for Euler against roughly 880,000 for Navier-Stokes — a ratio of about 176 to 1 for a problem OpenAI itself filed under “easier”.
The concurrent claim
OpenAI’s post also records that Levent Alpoge of Anthropic and Tristan Buckmaster of NYU had produced a resolution of the forced Euler problem using an internal Anthropic model. OpenAI says it offered a joint announcement, recognises their priority on forced Euler, and confirmed after an investigation that Buckmaster’s Codex prompts could not have influenced its system. That investigation followed a separate row in which mathematicians asked OpenAI to prove their unpublished work had not helped, which the math advisory group is now positioned to arbitrate.
The Declaration That Preceded the Math Advisory Group
The letter OpenAI cites by name in the math advisory group announcement is worth reading in the original rather than in summary, because the summaries have flattened it.
The core claim
Published on 11 September 2026 under the title “A Severe Misalignment of AI in Mathematics”, the declaration concedes the capability outright. The mathematical capabilities of these systems have improved “to the point that they can solve major outstanding problems in many fields of mathematics.” Its objection is to the use of those problems as benchmarks. “The goals of the AI companies and the goals of the mathematical community are severely misaligned.”
The argument about process
The signatories are not arguing that proofs are worthless. They argue that famous problems served as landmarks precisely because solving one required insight that could then be studied, discussed, simplified and eventually taught. “The mass production at faster and faster pace of ‘true/false’ statements could destroy fertile ground instead of breathing life into new ideas.”
The attribution complaint
The sharpest paragraph is about speed of publication. Solutions “are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions.”
How large the support base is
By 22 September 2026 the declaration carried 7,832 endorsers alongside its Fields Medallist signatories. Endorsements require an ORCID sign-in or a confirmed academic email address, so the number is a count of identified academics rather than open petition signatures. The grievance is not confined to OpenAI: Google has been running its own internal mathematics model on the same benchmark logic.
What the Math Advisory Group Is Not Allowed to Do
One sentence in OpenAI’s announcement does more work than the rest of it combined, and it defines the limits of the math advisory group precisely.
The pacing exclusion
“Importantly, the group will not be responsible for advising us on how to pace our internal progress on mathematics.” The declaration’s central request is that the pace of benchmark-driven problem solving be reconsidered. The math advisory group is constituted to advise on everything except that.
Advice versus authority
The math advisory group can assess significance, coordinate release timing, advise on standards, and advise on how OpenAI’s tools might support research and learning. It cannot slow anything down, and the IAS statement confirms it holds no decision-making power at any AI company. That is not a criticism of the members; it is the design of the arrangement, and the members agreed to it with the exclusion written down in public.
| What the declaration raises | Can the math advisory group address it |
|---|---|
| Results announced too fast for proper writeup | Yes, via dissemination advice |
| Prior work not cited, attribution unclear | Yes, via academic standards advice |
| Significance of a claimed result overstated | Yes, assessment is an explicit remit |
| Open problems used as marketing benchmarks | Only by public comment |
| The pace of internal mathematical progress | No, explicitly excluded |
| Loss of the human transmission chain | Indirectly, via tools and learning advice |
The leverage members do have
Three levers remain to the math advisory group. Members may publish unsolicited advice, they may comment publicly on OpenAI’s impact on mathematics, and they may resign. For a body of this composition, a coordinated resignation would be the loudest available signal, and the presence of three declaration signatories makes it a live possibility rather than a theoretical one.
The Twenty-Four Day Timeline
Laid out in order, the sequence explains the shape of the math advisory group announcement better than any of the individual documents does on its own.
From training run to governance structure
The model began training on 28 August. Eleven days later the Navier-Stokes proof was public. Three days after that the declaration appeared. Ten days after that, the math advisory group. The entire arc from the start of a training run to the creation of an external oversight structure took 24 days.
What the compression tells you
Ten days is fast for assembling nine senior mathematicians across seven institutions on two continents. Either the conversations predated the declaration, or the declaration accelerated something already in motion. OpenAI’s own wording — “we’re working with mathematicians who have established an independent mathematics advisory group” — places the initiative for the math advisory group with the mathematicians rather than with the company. That reading is consistent with three of the nine having already signed.
What the Math Advisory Group Means Outside Mathematics
Most readers of this will never touch a Millennium Prize problem. The pattern the math advisory group sets is still the one to watch.
The template being tested
A laboratory ships a capability, a professional community objects that the metric is wrong, and the laboratory responds with an unpaid advisory body whose remit excludes the thing the community asked about. This is the same shape as every other domain where model capability has outrun professional norms, and mathematics is simply the first field where the claims are formally checkable. A Lean proof either verifies or it does not. The same sequence has already played out with less checkable evidence: AI labs pressing ahead while insiders warn about control is the governance version of the argument the math advisory group now exists to have in public.
Why verifiability makes this the cleanest case
Mathematics is the rare discipline where an AI claim can be settled mechanically. OpenAI published a Lean formalisation alongside the writeup precisely so the result could be checked without trusting the company, the same posture it took when GPT-6 Astra crossed a cybersecurity threshold earlier in the month. In law, medicine or accounting there is no equivalent instrument. That is why independent testing of powerful AI models keeps returning as a demand across every sector — mathematics just has a head start on the tooling, and a math advisory group is what that head start looks like institutionally.
The governance question this leaves open
The honest reading is that the math advisory group is a genuine consultative body with no authority over the thing its own members have publicly objected to. Whether it is worth having depends entirely on whether those members use the public-comment lever. Organisations planning an AI strategy around frontier model claims should note how quickly this sequence ran, how little of it was subject to outside review before publication, and that the review body was constituted after the result shipped rather than before.
What to watch next
Three things would change the picture: publication of the list of 100-plus resolved problems, which OpenAI has not released; a public statement from the math advisory group itself rather than from OpenAI about it; and whether any further result is coordinated through the math advisory group before release rather than after. Until at least one of those happens, the arrangement is untested, and the fairest thing to say is that it has been announced rather than tried.
Frequently Asked Questions About the Math Advisory Group
Who runs the math advisory group?
The math advisory group is hosted at the Institute for Advanced Study in Princeton and controls its own membership. OpenAI does not appoint or pay its members, and the IAS has stated publicly that it holds no decision-making power at any AI company.
Has OpenAI published the 100 open problems?
No. The announcement states that the internal model has resolved more than 100 long-standing open problems across most areas of mathematics, but no list has been released. The Navier-Stokes proof and its Lean formalisation are the only published artefacts.
Did any math advisory group members sign the mathematicians’ letter?
Three of the nine did. Martin Hairer is an initial signatory on the declaration’s Fields Medallist list; Camillo De Lellis and Ravi Vakil both appear in its endorser table. Reports that only one math advisory group member had signed appear to have checked only one of the two lists.
Can the math advisory group stop or slow a release?
No. OpenAI’s announcement states explicitly that the body “will not be responsible for advising us on how to pace our internal progress on mathematics.” It can assess significance and advise on coordination and standards, and it can say publicly what it thinks.
Is the Navier-Stokes proof accepted?
It has been formally verified in Lean, which is a strong guarantee of internal correctness and the reason this claim is unlike most AI capability claims. Acceptance by the mathematical community is a separate and slower process, and it is exactly the process the declaration argues is being skipped.
Why does the signatory count differ between sources?
The archived DOI version of the declaration froze at 25 Fields Medallists on 11 September, three days after the Navier-Stokes proof appeared. The live page now lists 27, having added Vladimir Drinfeld and David Mumford. Coverage published from the archived version therefore reports the lower figure.
Is OpenAI claiming the $1 million prize?
No. The company stated it does not intend to claim the Millennium Prize for the result, framing the release as a report on the pace of AI progress rather than a prize submission.
References
Advisory Group on Mathematics and Artificial Intelligence — OpenAI
On the Navier-Stokes Millennium Prize Problem — OpenAI
A Severe Misalignment of AI in Mathematics — Declaration
Archived declaration, DOI 10.5281/zenodo.22737750 — Zenodo
A Severe Misalignment of AI in Mathematics — Terence Tao
OpenAI forms math advisory group as its AI resolves more than 100 open problems — TechCrunch
The Millennium Prize Problems — Clay Mathematics Institute
