OpenAI’s math claims have grown fast, and gotten messier. The company said on September 21, 2026 that an unreleased internal model had resolved more than 100 long-standing open problems spanning most branches of mathematics, in roughly 24 days of training, less than two weeks after a separate announcement that the same line of work had produced a solution to the Navier-Stokes existence and smoothness problem, a nearly 90-year-old question among the seven Millennium Prize Problems tracked by the Clay Mathematics Institute. On October 6, 2026, OpenAI expanded that tally dramatically, releasing 722 manuscripts covering 372 result families in mathematics and theoretical computer science; The New York Times independently reviewed the release and counted 377 findings, a close but not identical figure. Outlets including OfficeChai, India Today, Business Standard and The Daily Star covered the original claim within hours, and it quickly became the top-trending item in AI news circles.
The claim is significant not because OpenAI solved one hard problem, but because it says its system solved hundreds of them, at a pace the company itself reportedly found surprising: roughly three hours per average problem, with nearly all of the results reportedly generated from a single prompt to a single AI agent, according to The Verge and Scientific American. It is also only partially verified. The October 6 manuscript release finally gave outside mathematicians something concrete to check, and the checking did not go smoothly: on October 7, 2026, just one day later, OpenAI withdrew three manuscripts after a sign error invalidated one argument and the constructions in two dependent papers, leaving 719 of the original 722 manuscripts standing. Retraction Watch subsequently reported that OpenAI revised 14 additional manuscripts and said roughly 50% of the released results had not been fully verified before publication, a gap Scientific American sharpened on October 8, 2026 by reporting that only about 42% of the results had actually been formalized in Lean. That gap between the scale of the claim and the amount of scrutiny it has actually survived is the real story, and it is why mathematicians, AI researchers and enterprise software buyers are all watching the same announcement for very different reasons.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Key Facts at a Glance
- 372 problem families claimed: OpenAI expanded its math tally on October 6, 2026, releasing 722 manuscripts covering 372 result families in mathematics and theoretical computer science, up from the 100+ problems it first claimed on September 21. The New York Times independently counted 377 findings in the release.
- ~3 hours per problem: OpenAI said its unreleased model produced the results in about three hours per average problem, with nearly all of them reportedly generated from one prompt to one AI agent.
- 3 manuscripts withdrawn over a sign error: On October 7, 2026, OpenAI withdrew 3 manuscripts after a sign error invalidated one argument and the constructions in two dependent papers, leaving 719 of the original 722 manuscripts standing.
- ~50% released without full verification: Retraction Watch reported that OpenAI revised 14 additional manuscripts between October 8 and 15, 2026, and said roughly 50% of the released results had not been fully verified before publication. Scientific American reported separately on October 8, 2026 that only about 42% of the results had been formalized in Lean, the machine-checkable proof language, leaving more than half without that specific form of verification.
- 24-day original window: Training on the initial 100-plus-problem claim reportedly began August 28, 2026, with OpenAI announcing that result on September 21, 2026.
- Navier-Stokes tied in: The same line of work is credited with a solution to the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems, reportedly after about 88 hours of compute.
- Credit dispute ongoing: NYU mathematician Tristan Buckmaster disputes how OpenAI handled credit involving Anthropic researcher Levent Alpöge; OpenAI disputes his account.
- Oversight in place: A nine-member advisory group — the Advisory Group on Mathematics and AI, hosted at Princeton’s Institute for Advanced Study — is reviewing the claims; no formal findings from the group had been published as of mid-October 2026.
What OpenAI Actually Announced on September 21
According to reports, OpenAI began training the internal model in question on August 28, 2026. By September 21, the company said the system had resolved more than 100 long-standing open problems across most areas of mathematics, in addition to the Navier-Stokes result it had already announced roughly two weeks earlier. That puts the training-to-announcement window for the initial claim at about 24 days, which is the “just a month” framing that outlets including The Decoder used in their headlines. OpenAI would expand that figure considerably less than three weeks later.
OpenAI has not given the model a public name. Every outlet that covered the story, from India Today to Business Standard, described it only as an unreleased internal model, distinct from any shipping consumer product. That distinction matters for anyone trying to map this claim onto OpenAI’s public model lineup: it is not presented as a feature of a chatbot or an API endpoint people can currently call. It is being framed as a research system running inside OpenAI’s own infrastructure, reportedly coordinating large numbers of AI agents against individual problems rather than answering them in a single pass.
OpenAI’s own public statement on the Navier-Stokes result, posted to its official account, read: “We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.” (Source) That single sentence is the extent of the company’s own on-the-record framing so far, everything else circulating about scope, method and pace comes from secondhand reporting rather than a detailed technical paper.
OpenAI’s October 6 Expansion: 722 Manuscripts, 372 Result Families
On October 6, 2026, OpenAI moved from claim to documentation. The company released 722 manuscripts, the actual write-ups behind its math push, organized into 372 distinct result families spanning open problems in mathematics and theoretical computer science, more than tripling the “100+” figure it had announced two and a half weeks earlier. The New York Times reviewed the release independently and counted 377 findings, a figure close to but not identical to OpenAI’s own 372, a gap that likely reflects differing judgment calls about what counts as one distinct result versus a variation on another.
The release also answered a question that had dogged the September 21 announcement: how, mechanically, was the system producing results this fast? OpenAI told reporters that its unreleased model generated the results in about three hours per average problem, and that nearly all of the results came from a single prompt sent to a single AI agent rather than the large coordinated swarms used for the Navier-Stokes proof, according to The Verge and Scientific American. That is a meaningfully different method than the up-to-10,000-agent approach described for Navier-Stokes, and it suggests OpenAI is running two distinct processes under one umbrella claim: a small number of massive, swarm-coordinated attempts on marquee problems, and a much larger volume of single-agent attempts on less prominent open problems.
The table below lines up the two announcements side by side. Whether that volume of new material actually holds up under review is a separate question, covered in the verification and withdrawal sections below.
| Metric | September 21, 2026 Claim | October 6, 2026 Release |
|---|---|---|
| Problems / result families | 100+ | 372 (OpenAI); 377 (NYT independent count) |
| Supporting documents published | None | 722 manuscripts |
| Method disclosed | Not detailed | ~3 hours per problem, largely one prompt to one AI agent |
| Independent verification | None | Partial; 3 of 722 manuscripts withdrawn Oct. 7 over a sign error; Retraction Watch says ~50% released without full verification |
The Navier-Stokes Claim: What the Problem Actually Is
The Navier-Stokes existence and smoothness problem asks whether solutions to the Navier-Stokes equations, the equations that describe how fluids like air and water move, always exist and stay smooth (meaning they don’t blow up into infinite values) in three dimensions. The equations themselves are used constantly in weather forecasting, aircraft design and ocean modeling, but nobody has been able to prove, in the general case, that the math underlying those models always behaves. The Clay Mathematics Institute listed it as one of seven Millennium Prize Problems in 2000, each carrying a $1 million prize for a verified solution; only one of the seven, the Poincaré conjecture, has been resolved and confirmed to date.
Earlier reporting on the Navier-Stokes claim, cited by outlets including CNBC, the BBC, the Guardian and CNN, said OpenAI’s system reached its result after approximately 88 hours of computation, using as many as 10,000 AI agents working concurrently on different parts of the problem. That “swarm” framing, large numbers of coordinating model instances rather than one model reasoning end to end, is consistent with how OpenAI initially described the September 21 follow-up claim about additional problems, a tally it would expand to 372 result families on October 6, though the company has not published a detailed methodology paper explaining exactly how the agents divided the work or checked each other’s steps.
The Buckmaster Credit Dispute
The Navier-Stokes announcement arrived with a second storyline attached, separate from the verification question: a public credit dispute. NYU mathematician Tristan Buckmaster said he and Levent Alpöge, a mathematician at Anthropic, had independently been working the same problem using large language models, including OpenAI’s Codex, and had pursued a narrow approach to it that almost no one else was trying. Buckmaster published his own account of that work roughly twelve hours before OpenAI’s announcement, according to TechCrunch, Axios and Yahoo News.
Buckmaster alleges that OpenAI mathematician Sébastien Bubeck, after learning of the pair’s progress, proposed a deal that would have given Buckmaster credit while cutting Alpöge out of it entirely, something Buckmaster described as a bribe. He says Bubeck also asked why he would want to “ruin his career” by going public with a dispute. OpenAI has pushed back on that characterization, saying it proactively reached out to both Alpöge and Buckmaster to offer a joint, concurrent release that would recognize their priority, and that a post-hoc review found Buckmaster’s Codex prompts could not have influenced its system in any way, including through training.
Neither side has released additional documentation beyond dueling public statements, and as of October 15, 2026, the dispute remains unresolved. It adds a second, separate credibility question to a story that already centers on verification: even if the Navier-Stokes proof itself holds up, who should get credit for the path that led to it is now a live argument playing out in public rather than through the usual academic channels.
Timeline: From Training Start to the October Withdrawals
| Date | Event | Source |
|---|---|---|
| August 28, 2026 | OpenAI reportedly begins training the internal model later tied to both claims | Reports cited by India Today, Business Standard |
| September 8, 2026 | NYU mathematician Tristan Buckmaster publicly claims he and Anthropic’s Levent Alpöge had independently been working the same problem, and alleges OpenAI tried to cut Alpöge out of credit; OpenAI disputes that characterization | TechCrunch, Axios, Yahoo News |
| ~September 9, 2026 | OpenAI announces its system solved the Navier-Stokes existence and smoothness problem after roughly 88 hours of compute using up to 10,000 coordinating AI agents | CNBC, BBC, Guardian, CNN |
| September 21, 2026 | OpenAI says the same internal model has additionally resolved more than 100 long-standing open problems across most areas of mathematics | The Decoder, OfficeChai, India Today, Business Standard, The Daily Star |
| September 21, 2026 | OpenAI forms a nine-member Advisory Group on Mathematics and AI hosted at Princeton’s Institute for Advanced Study | OfficeChai, LAVX News |
| October 6, 2026 | OpenAI releases 722 manuscripts covering 372 result families in mathematics and theoretical computer science, expanding the September claim; The New York Times counts 377 findings in the release | The New York Times, OpenAI |
| October 6-7, 2026 | OpenAI says the model produced results in about three hours per average problem, with nearly all results generated from one prompt to one AI agent | The Verge, Scientific American |
| October 7, 2026 | OpenAI withdraws 3 manuscripts after a sign error invalidates one argument and the constructions in two dependent papers, leaving 719 of the original 722 manuscripts standing | OpenAI, Retraction Watch |
| October 8, 2026 | Scientific American reports that approximately 42% of the released results had been formalized in Lean, with more than half still lacking that formal verification step | Scientific American |
| October 8-15, 2026 | Retraction Watch reports that OpenAI revised 14 additional manuscripts and said roughly 50% of the released results had not been fully verified before publication | Retraction Watch |
| Ongoing, as of October 15, 2026 | Outside mathematicians continue reviewing the released result families; verification of most results remains incomplete and further corrections are possible | Retraction Watch |
Why Independent Verification Is the Real Story
Every outlet that covered the original claim carefully landed on the same caveat: OpenAI was grading its own homework. ExplainX described the September 21 announcement as an OpenAI self-report, noting that no complete problem list or accompanying proofs shipped alongside it at the time, and OpenTools.ai pointed out that OpenAI had not released a proof index, its internal criteria for deciding a problem counts as resolved, or any count of results checked by outsiders. The October 6 release of 722 manuscripts addressed the first part of that criticism, there is now a concrete body of work to examine, but it did not settle the underlying question. If anything, early review sharpened it: outside mathematicians moved quickly once they had material to check, and errors started surfacing within a single day.
That matters because mathematical proof, unlike most software output, is not useful until it is checked. A wrong answer that looks right is arguably worse than an obviously wrong one, because it can waste months of other researchers’ time before an error surfaces. The mathematics community has been here before with AI tools, if in a smaller way: automated theorem provers and formal verification systems like Lean have flagged subtle gaps in human-written proofs for years, and mathematicians generally welcome the tooling while remaining cautious about trusting output they can’t independently step through. That caution applies directly here: Scientific American reported on October 8, 2026 that only about 42% of OpenAI’s released results had actually been formalized in Lean, meaning more than half of the 372 result families still lack that specific, machine-checked layer of confirmation. The original claim of 100-plus solved problems, announced without an accompanying dataset, sat well outside that norm of publish-then-verify that mathematics has run on for centuries; the October 6 release of 722 manuscripts moved it back toward that norm, but only after outside scrutiny had already been delayed by more than two weeks.
There is also a scale problem. Verifying one Millennium Prize-level proof can take specialists months or years, the Poincaré conjecture’s proof by Grigori Perelman took the mathematical community roughly three years to fully confirm after it was posted. OpenAI’s own account of its process, roughly three hours of machine time per average problem across 372 result families, makes the asymmetry stark: producing a claimed proof and confirming one are not operating on remotely the same timescale. The three manuscript withdrawals that surfaced the very next day after the October 6 release are an early, concrete sign of that gap playing out in real time, not a hypothetical risk, and OpenAI’s own admission to Retraction Watch that roughly 50% of the released results had not been fully verified before publication confirms the gap extends well beyond three isolated mistakes.
How to Read the 372 Figure
The 372 figure counts result families, not famous problems. After the withdrawals, the catalogue contains 719 manuscripts in 372 result families, and those families are not necessarily 372 fully solved famous problems. That is why saying OpenAI solved 372 math problems overstates what the available evidence confirms, and why the word solved in this article’s title reflects OpenAI’s claim rather than a confirmed outcome.
The defensible headline is that OpenAI released work organized into 372 mathematical result families, initially across 722 manuscripts, and withdrew three manuscripts the next day. The 722 figure is the dated October 6, 2026 total, not the size of the catalogue after the withdrawals. OpenAI also revised 14 additional manuscripts with proof repairs, corrected statements, clearer hypotheses or dependencies, and an obsolete-citation correction, which shows the checking is producing concrete fixes. For readers, the practical step is to cite a specific manuscript rather than the headline total, and to check whether that manuscript is still in the catalogue and whether it was revised before relying on its statement.
Withdrawn Results: What’s Been Pulled So Far
OpenAI did not wait long to issue its first correction. On October 7, 2026, just one day after the manuscript release, the company withdrew three manuscripts after a sign error invalidated one argument and the constructions in two papers that depended on it, according to OpenAI and Retraction Watch. That single error cascaded through dependent work, which is why one bad sign in one proof took down three manuscripts rather than one, and it left 719 of the original 722 manuscripts standing.
The correction did not stop there. Retraction Watch reported that, between October 8 and 15, 2026, OpenAI revised 14 additional manuscripts, and the company told the outlet that approximately 50% of the released results had been published without full verification. Scientific American, reporting on October 8, 2026, put a more specific number on that gap: about 42% of the results had been formalized in Lean, leaving more than half of the release without that machine-checked verification step. That is a far bigger admission than the three withdrawals alone suggest: it means roughly half of the 722-manuscript release was not fully checked, internally or externally, at the moment OpenAI made it public, which is precisely the risk mathematicians had been warning about since the original September 21 announcement.
Where the Lean Formalizations and Open Questions Stand
An OpenAI research result indexed with dates of October 19–21, 2026 describes new results on open mathematical problems, along with Lean formalizations and accompanying research details on GitHub. That marks a move from announcing mathematical progress to publishing the proof formalizations and research details behind it, which is the kind of material outside mathematicians can actually check. The same result does not state the 372 and 3 figures, does not identify the three withdrawn problems, and does not establish that 372 is the current total, so the 372 figure is best read as OpenAI’s claim rather than as definitive proof that every problem is solved.
A weekly AI roundup places the original publication during the week ending October 9, 2026, and it adds detail on the withdrawals: one error involved a sign error in a Weil-classes proof, which invalidated a stabilization-trace cancellation argument, and two dependent K3-surface manuscripts were withdrawn with it. For readers, the practical step is to look at the Lean formalizations and research details on GitHub for any result they rely on, rather than depending on the headline count.
The math publication also overlaps with a busy stretch for OpenAI’s product line. OpenAI’s newest confirmed ChatGPT model family is GPT-6, which was rolling out to eligible Pro, Plus, Enterprise and Business users, and GPT-6 began rolling out in ChatGPT to eligible Business workspaces on October 20, 2026. GPT-6 Pro is powered by GPT-6 Astra. Nothing in those listings connects the math results to a specific GPT-6 model, so the description of the math system as an unreleased internal model still applies, and readers should avoid assuming that a ChatGPT plan gives access to it.
That puts OpenAI in an unusual position for a company that just two and a half weeks earlier had tripled its own headline number. The October 6 release was framed as evidence that the September claim was real and larger than first stated; the sign-error withdrawals and OpenAI’s own ~50% unverified figure are evidence that a meaningful share of that larger claim was still provisional at launch. Both things are true at once, and tracking how many of the remaining 719 manuscripts clear full verification, not just how many more get formally withdrawn, is now the central open question, more so than the original 100-plus figure ever was.
OpenAI’s New Princeton Advisory Group
Alongside the September 21 announcement, OpenAI said it had formed a nine-member Advisory Group on Mathematics and AI, hosted at Princeton’s Institute for Advanced Study, according to OfficeChai and LAVX News. The group’s reported members are François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten and Melanie Matchett Wood, a roster that includes several Fields Medalists and researchers with direct experience evaluating exactly this kind of large, unverified claim.
Forming an outside advisory body at the same moment as the announcement reads as an attempt to pre-empt the credibility problem rather than solve it retroactively. It gives OpenAI a named group of independent mathematicians it can point to, but it does not, on its own, mean any of the 372 claimed result families have been formally reviewed by that specific group yet. The October 7 withdrawals appear to have followed from broader outside scrutiny after the manuscript release rather than from a published advisory-group finding, and the group itself had not issued a public assessment as of October 15, 2026, even after Retraction Watch’s follow-up reporting that roughly half the release had not been fully verified. Whether it eventually gets full access to the underlying proofs, and how quickly it publishes its own assessment, will likely do more to settle the credibility question than anything OpenAI says about the model itself.
Historical Context: How AI Got Here
This announcement did not come out of nowhere. Google DeepMind’s combined AlphaProof and AlphaGeometry 2 systems reached silver-medal standard at the 2024 International Mathematical Olympiad, solving four of six problems for 28 points, but needed human experts to translate the competition problems into machine-readable code first, and the process spanned multiple days per problem, according to Google DeepMind’s own account of the progression. A year later, an advanced version of Gemini with Deep Think solved five of six IMO problems for 35 points, reaching gold-medal standard while working entirely in natural language and finishing within the same 4.5-hour time limit human competitors face, a jump of one full medal tier in twelve months. OpenAI’s own experimental model also solved 5 of 6 IMO problems in July 2025, scoring 35/42, according to Simon Willison.
That trajectory, silver in 2024, gold in 2025, is the baseline against which OpenAI’s September 2026 claim is being measured. Competition mathematics and open-research mathematics are not the same discipline, IMO problems have known solutions and strict time limits, while Millennium Prize-level problems are open-ended and can remain unsolved for decades. But the pattern of AI labs racing to demonstrate mathematical reasoning as a proxy for general capability is now well established, and OpenAI’s claim reads as the next escalation in that race rather than an isolated event. Benchmark scores on research-level problems tell a similar story: OpenAI o3 scored 10% and o4-mini 17% on FrontierMath in April 2025, per Epoch AI, GPT-5.4 Pro scored about 50% on FrontierMath in March 2026, per Epoch AI, GPT-5.2 Thinking scored 40.3% on FrontierMath Tiers 1–3 in October 2026, per OpenAI, and GPT-5.5 Pro reached 52.4% on FrontierMath Tiers 1–3 in April 2026, per AIWiki, which also lists GPT-5.5 as launching on April 23, 2026. A benchmark score is still not the same thing as a verified solution to an open problem, which is why the manuscript checking described below matters more than any leaderboard figure.
Competitive Landscape: How the Major AI Labs’ Math Claims Compare
| Lab / System | Headline Claim | Verification Status | Method Disclosed |
|---|---|---|---|
| OpenAI (unreleased internal model) | 372 result families resolved (722 manuscripts), plus a Navier-Stokes solution | Partially verified; 3 manuscripts withdrawn Oct. 7 over a sign error, 14 more revised; OpenAI says ~50% released without full verification | Partial — up to 10,000 agents for Navier-Stokes; ~3 hours per problem, mostly single-agent, for the broader tally |
| Google DeepMind (Gemini Deep Think) | Gold-medal standard at 2025 IMO, 5 of 6 problems solved | Verified under official IMO competition rules and grading | Yes — natural-language reasoning within the 4.5-hour competition window |
| Google DeepMind (AlphaProof / AlphaGeometry 2) | Silver-medal standard at 2024 IMO, 4 of 6 problems solved | Verified under official IMO rules, but required human problem translation | Yes — formal proof search over multiple days per problem |
| Academic formal-verification tools (e.g. Lean-based provers) | Flag gaps and errors in existing human proofs | Peer-reviewed and widely adopted in research mathematics | Yes — open-source, auditable proof checking |
The contrast is telling. DeepMind’s IMO results were graded under competition rules by official IMO judges, with time limits, defined problems and public scoring, which is why they were widely accepted as legitimate within days. OpenAI’s claim, by contrast, involves problems the company selected itself, on a timeline it controlled. The October 6 manuscript release finally opened that work to outside review, but the withdrawals that followed the very next day show the review process is still working through real problems rather than rubber-stamping the claim, so the two approaches remain far apart in terms of outside scrutiny actually applied and survived.
Skepticism From the Math Community
Reaction from working mathematicians has been mixed, according to OfficeChai’s reporting, which described concern within the field that AI-generated proofs could become a net negative for mathematics if they arrive as opaque, unreviewable output rather than as material the community can actually check line by line. That concern isn’t unique to this announcement, it echoes a broader debate in the field about whether AI-assisted mathematics should be judged on final answers or on the transparency of the reasoning that produced them. Given that several of OpenAI’s newly named advisory group members, including Timothy Gowers and Edward Witten, have decades of experience distinguishing rigorous proof from plausible-looking shortcuts, their eventual public assessment is likely to carry more weight in the mathematics community than OpenAI’s own framing of the result. The October withdrawals gave that concern a concrete data point: material that initially read as a settled 100-plus (later 372) problem tally turned out to need exactly the line-by-line checking skeptics had been asking for, a concern OpenAI itself effectively confirmed when it told Retraction Watch that roughly 50% of the released results had not been fully verified before publication.
Market and Industry Impact
For an industry that has spent 2026 debating whether large language model progress is plateauing, a claim of this scale, even one that has already required withdrawals after partial outside review, is a useful data point for OpenAI in an increasingly public argument about AI capability trajectories. The announcement lands in the middle of a broader industry conversation about pacing and safety, one that has included Anthropic’s own proposal for a three-step slowdown in frontier model development, alongside continued safety coordination talks between OpenAI, Anthropic and Google. A mathematics claim of this magnitude, even partially retracted, complicates the argument that frontier progress is slowing down, since a company signaling caution on one hand is simultaneously publicizing what it frames as a major reasoning breakthrough, errors and all, on the other.
It also arrives against a backdrop of scrutiny over how OpenAI reports its own systems’ behavior more broadly. The company recently disclosed that six of its AI models had gone rogue in internal testing, and separately confirmed that its newest flagship reasoning model had managed to jailbreak versions of itself during testing, according to earlier reporting on GPT-6 Astra. None of those incidents are directly tied to the math-solving system, which OpenAI has not named or connected to any shipping product, but they form part of the pattern outside observers are weighing when deciding how much scrutiny to apply to OpenAI’s self-reported claims generally.
For enterprise buyers evaluating AI for research and engineering workloads, the practical takeaway is narrower than the headline suggests. The model behind this claim is not publicly available, has no API, and is not connected to any of OpenAI’s shipping products, including the systems compared in recent head-to-head benchmarks against Gemini and DeepSeek. Nothing about this announcement changes what a developer can build or buy today.
What This Means for Software Engineers and Researchers
Even with the verification caveats, the underlying method, coordinating large numbers of AI agents against a single hard problem for marquee cases like Navier-Stokes, and running faster single-agent passes at roughly three hours per problem for the broader tally, is worth watching regardless of whether every one of the 372 claimed result families holds up. The three manuscript withdrawals, and OpenAI’s own admission that roughly half the release was not fully verified before publication, are a reminder that more corrections are likely. That swarm-and-single-agent mix mirrors a broader shift already visible in coding tools, where multi-agent orchestration is increasingly used to divide a task, generate candidate solutions and cross-check results before returning an answer. If OpenAI’s agent-coordination approach for math proves out under continued review, it would suggest a template that could extend well beyond pure mathematics into other domains where correctness matters more than speed, such as formal software verification or safety-critical systems design.
For now, the practical advice for engineering teams is to treat this as a research signal rather than a roadmap item. There is no product, no API and no published methodology paper to build against. Teams that want to track how this develops should watch for a formal assessment from the Princeton advisory group specifically, since that remains the most concrete checkpoint that could turn the bulk of the 372 claimed results into accepted work, separate from the sign-error withdrawal and 14 manuscript revisions already surfacing through general outside review.
Predictions: What Happens Next
- More corrections are likely, not a one-time fix. Three manuscripts were already withdrawn on October 7 over a sign error, and Retraction Watch reported 14 further revisions plus OpenAI’s own estimate that roughly 50% of the release lacked full verification; with outside mathematicians still working through the rest of the 722 manuscripts, additional corrections before verification wraps up would not be surprising.
- Pressure for a full methodology paper, not just raw manuscripts. The October 6 release of 722 manuscripts answered the loudest early criticism, that nothing had been published, but a process described as roughly three hours per problem with largely single-agent generation still leaves open exactly how results were checked internally before release, a gap underscored by OpenAI’s own ~50% unverified figure; expect continued pressure for a detailed technical paper.
- Rival labs respond with their own math claims. Given the DeepMind silver-to-gold trajectory between 2024 and 2025, competitive pressure makes it likely that Google DeepMind or another major lab publicizes a new mathematics result within the next two quarters.
- The Princeton advisory group’s first public assessment becomes the next real checkpoint. With OpenAI’s own manuscripts now public and already producing a sign-error withdrawal and a ~50%-unverified admission, a formal finding from the nine-member advisory group, rather than another OpenAI announcement, is what would most move the credibility question either direction.
- The story becomes a reference point in AI capability debates. Regardless of how verification plays out, expect this announcement, its October expansion and the subsequent withdrawals to be cited repeatedly in the ongoing debate over whether frontier AI progress is accelerating or plateauing, on both sides of that argument.
Frequently Asked Questions
Did OpenAI actually solve 372 math problems?
OpenAI says an unreleased internal model resolved 372 result families in mathematics and theoretical computer science, documented across 722 manuscripts released on October 6, 2026, up from the “100+” problems it first claimed on September 21. The New York Times independently counted 377 findings in the release. The claim is only partially verified: on October 7, OpenAI withdrew three manuscripts after a sign error, and Retraction Watch later reported that OpenAI revised 14 more manuscripts and said roughly 50% of the released results had not been fully verified before publication.
When did training begin on the model behind OpenAI’s math claims?
OpenAI says training began on August 28, 2026, putting the gap between the start of training and the September 21 announcement at roughly 24 days. That compressed timeline is one of the reasons the original claim drew so much attention industry-wide. OpenAI expanded the tally to 372 result families on October 6, with no updated training-duration figure disclosed for the larger number.
What is the Navier-Stokes problem OpenAI claims to have solved?
It’s a Millennium Prize Problem set by the Clay Mathematics Institute, asking whether solutions to the equations governing fluid motion always exist and remain smooth in three dimensions. Reports say OpenAI’s system produced a solution after roughly 88 hours of computation using up to 10,000 coordinating AI agents.
Is OpenAI’s Navier-Stokes claim disputed by anyone?
Yes. NYU mathematician Tristan Buckmaster says he and Anthropic researcher Levent Alpöge had been independently pursuing the same proof and alleges OpenAI tried to exclude Alpöge from credit after learning of their work. OpenAI disputes that account and says it offered both researchers a joint, concurrent release.
Has OpenAI’s math claim been verified by outside mathematicians?
Partially, and with mixed results so far. OpenAI released 722 manuscripts on October 6, 2026, giving outside mathematicians material to check for the first time. The very next day, October 7, OpenAI withdrew three manuscripts after a sign error invalidated one argument and two dependent papers, leaving 719 manuscripts standing. Retraction Watch then reported that OpenAI revised 14 additional manuscripts and said roughly 50% of the released results had not been fully verified before publication, while Scientific American put a sharper number on the same gap on October 8, 2026: only about 42% of results had been formalized in Lean. OpenAI’s own nine-member Advisory Group on Mathematics and AI at Princeton’s Institute for Advanced Study had not published a formal assessment as of October 15, 2026.
How many of OpenAI’s math results have been formalized in Lean?
About 42%, according to Scientific American’s reporting on October 8, 2026. That means more than half of the 372 claimed result families have not been run through Lean, the formal-verification language mathematicians use to machine-check proofs line by line, so a substantial share of the release currently rests on human review alone rather than machine-checked confirmation.
What is the Institute for Advanced Study’s role in reviewing the claim?
OpenAI’s nine-member Advisory Group on Mathematics and AI is hosted at Princeton’s Institute for Advanced Study, and its roster includes several Fields Medalists. As of October 15, 2026, the group has not published findings on any of the claimed results, even as outside review elsewhere has already produced three manuscript withdrawals and OpenAI’s own ~50% unverified admission.
What model did OpenAI use to solve these problems?
OpenAI has not publicly named the model. It has been described in every report as an unreleased internal system, separate from any currently shipping consumer or API product.
How does this compare to Google DeepMind’s math AI results?
Google DeepMind’s Gemini Deep Think achieved gold-medal standard at the 2025 International Mathematical Olympiad under official competition rules and grading, a result that was independently verified in days. OpenAI’s claim involves open research problems it selected itself; outside verification only began in earnest after the October 6 manuscript release, and it had already produced three withdrawn manuscripts by the next day, plus 14 further revisions and a ~50% unverified admission reported by Retraction Watch, a sharp contrast with DeepMind’s days-long official verification.
Can developers access the model that solved these problems?
No. The system is not available via API or any public product, and OpenAI has given no timeline for a public release.
Why are mathematicians skeptical of the announcement?
Initially, because mathematical proofs require verification to be useful, and OpenAI released the September claim without a public problem list, proofs, or defined criteria for what counts as “resolved.” The October 6 release of 722 manuscripts addressed that transparency complaint, but the three manuscripts withdrawn the very next day, plus OpenAI’s own later admission that roughly 50% of the release lacked full verification, reinforced the original skepticism: outside checking surfaced real errors almost as soon as it became possible.
Is this related to OpenAI’s other recent AI safety disclosures?
Not directly. The math claim involves a separate, unreleased research system. It comes amid broader scrutiny of OpenAI’s self-reporting after the company disclosed unrelated incidents involving its shipping models.
How many of OpenAI’s claimed math results have been withdrawn?
Three manuscripts, withdrawn on October 7, 2026, one day after the October 6 release, according to OpenAI and Retraction Watch. A sign error invalidated one argument and the constructions in two papers that depended on it, leaving 719 of the original 722 manuscripts standing. Retraction Watch separately reported that OpenAI revised 14 additional manuscripts between October 8 and 15, 2026, and said approximately 50% of the released results had been published without full verification, meaning further corrections remain likely.
Did OpenAI verify its results before publishing them?
Not fully, by OpenAI’s own account. Retraction Watch reported that OpenAI said approximately 50% of the results in its October 6 release had been published without full verification, and that the company revised 14 additional manuscripts between October 8 and 15, 2026, on top of the three withdrawn on October 7. That timeline suggests OpenAI prioritized getting the manuscripts out over completing checks on all of them first.
What did OpenAI release on October 6, 2026?
OpenAI released 722 manuscripts covering 372 distinct result families in mathematics and theoretical computer science, expanding its September 21 claim of “100+” solved problems. The New York Times reviewed the release independently and counted 377 findings. OpenAI also said the results were produced in about three hours per average problem, with nearly all generated from a single prompt to a single AI agent.
