OpenAI told the public on September 21, 2026 that an internal model had resolved more than 100 long-standing open problems spread across most areas of mathematics, and that the same system had also produced a claimed resolution of the Navier-Stokes existence and smoothness problem, one of the seven Clay Mathematics Institute Millennium Prize problems. The company did not name the model, did not publish an itemized list of the problems, and did not release the underlying proofs for outside review. Alongside the claim, OpenAI announced a new Advisory Group on Mathematics and Artificial Intelligence, a panel tasked with reviewing emerging mathematical results from AI systems before they get treated as settled fact.
That combination, a huge claim paired with an admission that outside checking hasn’t happened yet, is the actual news here. Training reportedly began on August 28, 2026, roughly three and a half weeks before the announcement. Multiple outlets, including TheDecoder, picked up the story within hours, but nearly every specific number beyond the “100-plus problems” figure and the Navier-Stokes claim remains unconfirmed by OpenAI itself.
What OpenAI Announced on September 21
The announcement itself was short on specifics. OpenAI said its model had resolved “more than 100 long-standing open problems” across most branches of mathematics, and separately flagged the Navier-Stokes result as a standout. It did not disclose which model produced the results, whether it was a variant of a publicly available system or a purpose-built research model, or what compute budget was used beyond the late-August start date. No architecture details, parameter counts, or training-duration figures were included in the cited materials.
That silence matters because OpenAI has made large math claims before. The company’s earlier Navier-Stokes proof announcement generated its own wave of coverage, and the September 21 update reads, in part, as a follow-up to that earlier controversy rather than a wholly new event. What changed is the addition of a formal review structure. Instead of simply repeating the claim, OpenAI is now telling reporters and mathematicians that a dedicated panel will vet results going forward, which is itself an implicit acknowledgment that the initial claims arrived without that scrutiny.
The Claim: 100-Plus Problems and a Reported Navier-Stokes Result
Start with what OpenAI actually said, stripped of amplification. The company reported that its model resolved more than 100 open mathematical problems. It separately said the same system produced a result for the Navier-Stokes Millennium Prize problem, a question about whether smooth, well-behaved solutions to the Navier-Stokes equations always exist in three dimensions, first posed formally by the Clay Mathematics Institute in 2000 with a $1 million prize attached.
Everything past that headline is murkier. Reports describing a specific number of participating AI agents, a precise training duration measured in hours, or a lengthy formal paper with machine-checked verification circulated widely online in the days after the announcement, but none of those specifics appear in OpenAI’s own cited statement. Treat those figures as unconfirmed until OpenAI or a named source attaches them directly to the record. The company’s own language leans on “the announcement” and general claims rather than a documented, line-by-line account of what was solved and how.
This is a familiar pattern in AI research announcements: a splashy headline number, followed by a slower, messier process of outside groups trying to reconstruct exactly what happened. It happened with early GPT-4 benchmark claims, it happened around the GPT-6 Astra training run disclosures, and it’s happening again here.
Why Independent Verification Hasn’t Happened Yet
Mathematical proofs, unlike benchmark scores, aren’t verified by running a test suite. A claimed proof of a Millennium Prize problem has to survive review from specialists who work in that exact sub-field, often over months or years, checking every logical step and every edge case. The Poincaré conjecture, the only Millennium Prize problem solved to date, took the mathematics community roughly three years of checking after Grigori Perelman posted his papers before the Clay Institute formally recognized the result in 2010.
OpenAI’s cited announcement does not provide an itemized list of the 100-plus problems, does not release the proofs themselves, and does not point to independent verification of any single result. That’s not necessarily evidence the claim is false, but it does mean the story, as of September 22, 2026, is “a company says it has extraordinary results,” not “extraordinary results have been confirmed.” Reporters covering the story, including TheDecoder, have been careful to frame it as an OpenAI claim rather than an established mathematical fact, and that distinction is the whole story.
Formal verification tools such as Lean, the proof assistant increasingly used to machine-check complex mathematics, offer one possible path to faster confirmation. Mathematicians including Fields Medalist Terence Tao have written publicly in recent years about using Lean to check both human and AI-assisted proofs, arguing that machine verification can shortcut some of the social trust problem that slows down peer review. Whether OpenAI’s claimed results were run through that kind of formal checking, or through anything resembling it, is not stated in the company’s own announcement.
Inside the Advisory Group on Mathematics and Artificial Intelligence
The advisory group is the part of this story that’s actually new. OpenAI describes its purpose as reviewing emerging mathematical results produced by its systems, assessing how significant those results actually are, and advising on how and when to share them publicly. In effect, it’s an internal check on the exact kind of claim OpenAI just made days before announcing the group’s existence.
OpenAI has not published a full roster of the group’s members, its review timeline, or whether it will have authority to block a public announcement it considers premature. Those are the questions that will determine whether the group is a substantive check on hype or a public-relations gesture attached to a claim that had already gone out the door. A review board announced after the claim it’s supposed to be reviewing carries a different weight than one stood up in advance.
What the Group Is Expected to Do
Based on OpenAI’s own description, the group’s mandate covers three things: reviewing results before or shortly after they’re claimed, judging how significant a given result actually is against the existing mathematical literature, and shaping how OpenAI communicates about those results to the public and the research community. None of those functions require the group to independently reproduce a proof from scratch, which means it could plausibly sign off on framing and communication without settling whether the underlying mathematics holds up.
Training Timeline: Confirmed Facts vs. Open Questions
Two dates anchor this story. Training reportedly started August 28, 2026. The public announcement landed September 21, 2026, about 24 days later. That’s the full confirmed timeline. Claims about a specific training duration measured in hours, a specific count of AI agents deployed in parallel, or a specific page count for an accompanying paper are not confirmed by OpenAI’s own cited statement and should be treated as unverified until OpenAI attaches them to the record directly.
That gap between what’s confirmed and what’s circulating matters for anyone trying to assess the claim honestly. A 24-day window between training start and public announcement is short by the standards of traditional mathematical publication, where results typically go through months of internal review, preprint circulation on repositories like arXiv, and referee comments before anyone calls them settled.
OpenAI’s Claim vs. What’s Actually Confirmed
| Element | What OpenAI Reported | Verification Status |
|---|---|---|
| Training start date | August 28, 2026 | Stated by OpenAI |
| Announcement date | September 21, 2026 | Confirmed, widely reported |
| Problems resolved | “More than 100” across most math areas | OpenAI claim, no itemized list published |
| Navier-Stokes result | Claimed resolution of the Millennium Prize problem | Not peer-reviewed or independently confirmed |
| Model name / architecture | Not disclosed | Unconfirmed |
| Training duration (hours/days) | Circulating in secondary reports | Not confirmed by OpenAI’s cited statement |
| Number of AI agents used | Circulating in secondary reports | Not confirmed by OpenAI’s cited statement |
| Advisory Group formation | Announced September 21, 2026 | Confirmed by OpenAI |
How This Compares to Prior AI Math Milestones
OpenAI isn’t the only lab chasing headline-grabbing math results. Google DeepMind’s AlphaProof and AlphaGeometry systems were reported to have reached gold-medal-equivalent performance at the International Mathematical Olympiad in the summer of 2025, a result the company presented with more procedural transparency than what OpenAI has offered here, including coordination with competition organizers on grading. OpenAI separately said, around the same period, that one of its own reasoning models had reached a comparable Olympiad-level score under research conditions.
The difference between an Olympiad problem and a Millennium Prize problem is enormous. Olympiad problems, however hard, have known solutions and clear grading rubrics built for teenagers to solve in hours. Millennium Prize problems have resisted the entire mathematical community for decades; only one of the original seven has ever been solved. Comparing “gold medal at a math competition” to “resolved a Millennium Prize problem” is comparing two different categories of difficulty, and OpenAI’s September 21 claim sits in the second, far more consequential, category, without the kind of verification process that would normally accompany a claim at that level.
Anthropic has taken a different public posture on AI-driven research, framing Claude’s contribution to its own internal R&D work in incremental, measured terms rather than announcing solved century-old problems. That contrast in communication style has become part of the broader industry conversation about how AI labs should talk about their own results, particularly as the same companies whose CEOs have publicly floated slowing down the pace of AI development continue to compete on who can claim the most dramatic capability milestone.
AI and Mathematics: A Short Timeline of Milestones
| Period | Milestone | Verification Path |
|---|---|---|
| 2000 | Clay Mathematics Institute names the seven Millennium Prize problems | Formal institutional process, peer-reviewed criteria |
| 2002-2003 | Grigori Perelman posts a proof of the Poincaré conjecture | Roughly three years of community review before formal recognition in 2010 |
| Summer 2025 | Google DeepMind and OpenAI both report Olympiad-level math performance from AI systems | Partial coordination with competition graders; not peer-reviewed research proofs |
| August 28, 2026 | OpenAI reportedly begins the training run behind its math claim | Date stated by OpenAI |
| September 21, 2026 | OpenAI announces 100+ solved problems, a claimed Navier-Stokes result, and forms the Advisory Group on Mathematics and AI | Announcement made; independent verification pending |
Market and Industry Reaction
The reaction across the AI industry split along familiar lines. Some coverage, including pickup from TheDecoder, treated the announcement as a genuine milestone worth tracking closely precisely because of who made it and how quickly it was reported across multiple outlets. Other commentators focused on the gap between the scale of the claim and the thinness of the supporting documentation, noting that a result of this magnitude, if fully verified, would likely be the single biggest mathematics story of the decade, not a single-paragraph announcement bundled with a new oversight committee.
For OpenAI specifically, the timing lands during a stretch in which the company has been racing to demonstrate that its frontier models, including the systems behind its large-scale training runs, are producing research-grade output rather than just better chatbots. Co-founder Greg Brockman has previously framed OpenAI’s large training efforts in sweeping, historic terms, and this math announcement fits that same pattern of positioning the company’s models as engines of scientific discovery, not just consumer products.
What Mathematicians Are Watching For
The mathematics community’s response to AI-generated proofs has generally converged on a few concrete asks: publish the actual proofs, not just a count of problems solved; make the work available for formal verification through tools like Lean; and let named specialists in the relevant sub-fields review the material before treating it as settled. None of those steps have happened yet for OpenAI’s September 21 claim. Until they do, mathematicians are likely to treat the announcement the same way they treated the company’s earlier Navier-Stokes claim: as an interesting data point, not a solved problem.
There’s also a practical wrinkle specific to the Navier-Stokes problem. The Clay Mathematics Institute’s rules for awarding a Millennium Prize require publication in a qualifying peer-reviewed journal followed by two years of general acceptance in the mathematical community before a prize is even considered. Even in the best case for OpenAI, where every claimed step held up perfectly, the institutional clock on an actual prize award would still run for years, not weeks.
The Business Stakes Behind the Claim
There’s a commercial angle here too. OpenAI, like its rivals, is under constant pressure to show that its enormous compute spending translates into capability gains that matter beyond chatbot benchmarks. A model that can meaningfully accelerate mathematical research, if the claim holds up, would be a powerful argument for continued investment in frontier-scale training runs. That’s part of why skepticism about the specifics doesn’t necessarily mean skepticism about the underlying possibility; AI-assisted mathematics is a real and growing field, and even partial progress toward automating parts of proof discovery would be commercially significant for a company competing directly with rivals making similarly bold capability claims.
At the same time, a claim that doesn’t survive scrutiny carries its own cost. OpenAI’s credibility on technical claims has already been tested once this year around the earlier Navier-Stokes announcement, and a second high-profile math claim that also fails to produce verifiable documentation would compound doubts about how the company communicates its research, right as it’s asking outside mathematicians to trust a review process it just created.
Predictions: Where This Story Goes Next
- Expect OpenAI to publish at least a partial list of the claimed 100-plus problems within weeks, likely starting with the least controversial, easiest-to-verify results first.
- The Navier-Stokes claim specifically will draw the most scrutiny and the slowest resolution; expect specialist mathematicians to spend months, not days, assessing any released materials.
- The Advisory Group on Mathematics and Artificial Intelligence will likely publish its membership and review criteria before it publishes any actual verdict on OpenAI’s claims, since establishing legitimacy for the process will come before using it.
- Competitors including Google DeepMind and Anthropic are likely to respond with their own framing of AI-assisted mathematics work, either by highlighting more conservative, already-verified results or by announcing comparable big claims of their own.
- If no itemized, checkable proof set appears within a few months, expect the September 21 announcement to be remembered the way the earlier Navier-Stokes claim has been: as a marketing moment rather than a confirmed mathematical result.
FAQ
Did OpenAI actually solve the Navier-Stokes Millennium Prize problem?
OpenAI said its model produced a result resolving the Navier-Stokes existence and smoothness problem. That claim has not been independently verified, peer-reviewed, or confirmed through the Clay Mathematics Institute’s award process as of September 22, 2026.
What is the Navier-Stokes existence and smoothness problem?
It’s one of seven Millennium Prize problems named by the Clay Mathematics Institute in 2000, asking whether smooth, physically reasonable solutions to the Navier-Stokes equations always exist in three dimensions. A full explanation of the problem’s mathematical background is available on Wikipedia’s entry on the topic.
What is OpenAI’s Advisory Group on Mathematics and Artificial Intelligence?
It’s a panel OpenAI announced on September 21, 2026, tasked with reviewing emerging mathematical results from its AI systems, assessing their significance, and advising on how those results are shared publicly.
Has any mathematician independently verified OpenAI’s claims?
Not as of this writing. OpenAI’s own cited announcement does not include an itemized list of the 100-plus problems or the underlying proofs, so outside specialists have not had material to review yet.
Which model did OpenAI use to produce these results?
OpenAI did not publicly name a specific model or disclose its architecture in the cited announcement. Reports linking the claim to a specific named system are unconfirmed.
How does this compare to Google DeepMind’s AI math work?
DeepMind’s AlphaProof and AlphaGeometry systems were reported to reach gold-medal-equivalent results at the 2025 International Mathematical Olympiad, with some coordination with competition graders. That’s a different, far less difficult category than resolving a Millennium Prize problem, and DeepMind’s process involved more public procedural detail than OpenAI has released for its September 21 claim.
Could OpenAI actually collect the $1 million Millennium Prize?
Not quickly, even in the best case for the claim. The Clay Mathematics Institute requires publication in a qualifying peer-reviewed journal and roughly two years of general acceptance in the mathematical community before a prize is considered.
What happens next?
Watch for whether OpenAI publishes an itemized list of the claimed problems and the underlying proofs, whether the new advisory group releases its membership and review criteria, and whether any named mathematician goes on record confirming or disputing individual results.
