OpenAI spent October rolling out one of the largest mathematics research drops any AI lab has ever published: 722 manuscripts spanning 372 result families, covering open problems across mathematics and theoretical computer science, according to OpenAI’s own announcement and coverage from GIGAZINE and The Indian Express. That total didn’t hold for even 24 hours: on October 7, 2026, OpenAI confirmed a sign error had invalidated one proof and two manuscripts that depended on it, pulling the standing catalog down to 719 papers. But the bigger story isn’t the page count or the correction. It’s the fight breaking out over whether any of it can be trusted without independent verification, and why OpenAI quietly formed an outside panel of mathematicians, including Terence Tao, just weeks before the release landed. Tao himself said on October 15, 2026 that he had reviewed approximately 10 of the released papers, which is a long way from vetting the full collection, and the correctness of all 719 manuscripts has still not been established.
The tension sits at the center of a credibility test for the entire AI industry. OpenAI has spent 2026 claiming its internal frontier model cracked longstanding problems in group theory, Ramsey theory, and extremal graph theory, and in September reportedly proposed a resolution to the Navier-Stokes existence and smoothness problem, one of the seven Clay Mathematics Institute Millennium Prize problems. Mathematicians outside the company are not convinced the claims hold up to the standard of proof the field has used for centuries, and that gap between “a model generated this” and “a mathematician verified this” is now the whole story.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI Actually Announced on October 6
The Headline Numbers
OpenAI said it was sharing “a broad range of new mathematical results” generated by an internal frontier model, publishing the work on GitHub rather than through a single splashy announcement, per OpenAI’s own release notes. The scale is what distinguishes this drop from prior AI-math moments: 722 manuscripts, organized into 372 result families, reported by GIGAZINE and The Indian Express as the headline figures. The Indian Express also reported that the manuscripts cover solutions to, or meaningful progress on, hundreds of open problems spanning both pure mathematics and theoretical computer science. Other reporting puts the claims at full or partial solutions to more than 370 open problems, including claimed progress on three Millennium Prize problems and a claimed proof of the Unique Games Conjecture, and the public repository includes papers, proof artifacts, Lean formalizations, and 10 abridged reasoning summaries.
How OpenAI Tested the Model
According to The Indian Express, OpenAI tested its internal frontier model against roughly 4,000 mathematical problems, with the average published result requiring computing power roughly equivalent to about three hours of ChatGPT Pro reasoning. That detail matters for understanding the economics of the project: this wasn’t a single breakthrough moment, it was an industrial-scale sweep across thousands of problems, with the strongest few hundred results curated for publication.
Readers tracking the earlier chapter of this story, including the initial 722-manuscript release from OpenAI’s unreleased model, already know the raw numbers. What’s changed in the days since is the response from the mathematics community, and it is not uniformly celebratory.
Why OpenAI Called In Outside Mathematicians
On September 21, 2026, OpenAI announced the Advisory Group on Mathematics and Artificial Intelligence, a nine-member panel hosted at Princeton’s Institute for Advanced Study, according to OpenAI’s own description of the group. Fields Medalist Terence Tao, who has become one of the most prominent voices scrutinizing AI-generated mathematics, explained the group’s purpose directly on his blog. “We are currently facing the very specific challenge of advising OpenAI on how to coordinate the release of a large number of significant results in mathematics that they report have been produced by their internal model,” Tao wrote on his blog when the group’s formation became public.
What the Advisory Group Actually Does
That single sentence does a lot of work. It confirms OpenAI itself describes the manuscripts as results the company “reports” were produced by its model, not results the advisory group has independently blessed as correct. It also confirms the advisory group’s job is specifically about coordinating a release, which is a different task than certifying mathematical truth. Business Standard reported that the advisory group’s creation traces back to OpenAI’s claim, made public in September, that its internal model had solved or made substantial progress on more than 100 longstanding problems after training that began on August 28, less than a month earlier.
The Pace That Prompted Scrutiny
The pace of that claim, more than 100 problems touched in under four weeks, is part of why outside mathematicians wanted a seat at the table before the full manuscript set went public. Speed at that scale is exactly the kind of claim that invites scrutiny in a field where a single proof can take a career to produce and years to check.
Who Sits on the Advisory Group
OpenAI’s description of the panel puts its total membership at 9 unpaid mathematicians. Tao is the most visible member because of his public blog posts explaining the group’s mandate, but the roster also includes Timothy Gowers and Martin Hairer, alongside theoretical physicist Edward Witten. None of the nine are paid for the work, which is itself a signal: OpenAI is leaning on volunteer goodwill from senior figures in the field rather than contracting reviewers, a structure that caps how much review bandwidth the group can realistically absorb no matter how many manuscripts OpenAI publishes.
The Open Letter That Triggered the Pushback
Before the advisory group existed, a group of mathematicians had already gone public with concerns. OpenAI’s own page describing the advisory group references the letter directly: “In a recent open letter ‘A Severe Misalignment of AI in Mathematics,’ mathematicians raise concerns about the negative externalities of solving open problems as a benchmark for new AI systems,” according to OpenAI’s advisory group announcement.
The “Negative Externalities” Argument
That phrase, “negative externalities of solving open problems as a benchmark,” is the crux of the mathematics community’s objection. The worry isn’t necessarily that the model is wrong. It’s that turning open problems into a scoreboard for AI labs changes the incentives around how results get produced, checked, and credited, potentially faster than the field’s verification norms can keep up. A wrong answer dressed up as a solved Erdős problem can cost other researchers months chasing down whether a published “breakthrough” is sound.
The Navier-Stokes Claim Still Isn’t “Solved”
The clearest test case for OpenAI’s verification problem is the Navier-Stokes existence and smoothness problem, one of the seven Clay Mathematics Institute Millennium Prize questions, each carrying a $1 million prize for a verified solution. Galaxy Research summarized the claim plainly: “The frontier lab said it had solved the Navier-Stokes (NS) equation, one of the seven Millennium Prize questions,” in its coverage of the dispute.
What the Official Record Shows
But the official record doesn’t match the announcement’s tone. “At the time of writing, the proof it generated was still being verified by external mathematicians, and the Clay Mathematics Institutes marked the question as ‘active’ instead of ‘solved,'” Galaxy Research reported. The Clay Mathematics Institute’s own list of Millennium Prize problems is the authoritative record of which of the seven questions have been formally resolved, and as of this reporting, Navier-Stokes is not listed among them.
How the Disputed Proof Was Produced
Earlier reporting described how the claimed result was produced: the BBC and CNN both reported that OpenAI used roughly 10,000 AI agents working for about 88 hours to generate the proposed Navier-Stokes result. That’s a genuinely new mode of mathematical production, agent swarms grinding on a problem in parallel rather than one researcher or even one model working linearly. Whether that production method yields something a human referee will eventually certify as a real proof is still an open question, and the BBC’s own coverage flagged that the claim had not been independently verified or accepted by the Clay Mathematics Institute at the time of its reporting.
How OpenAI Says It’s Checking Its Own Work
OpenAI isn’t ignoring the verification gap. According to Nature’s coverage of the claim, “OpenAI said that its breakthrough, which cost several million US dollars, has been validated using an automated verification method that is becoming the standard for rigour in mathematics,” in a report published by the scientific journal. That automated-verification detail is significant on two fronts: it puts a cost figure on the Navier-Stokes effort for the first time in reputable coverage, and it signals OpenAI is leaning on formal proof-checking software rather than asking the world to simply trust a language model’s output.
Automated Proof-Checking Explained
Automated proof checkers, tools that can mechanically confirm a sequence of logical steps follows valid inference rules, have existed in mathematics for decades and are the backbone of projects like formalized mathematical libraries. Pairing a frontier model’s output with that kind of machine-checkable verification is a real step toward trustworthiness. It is not, however, the same as a peer-reviewed publication or a field-wide consensus that a Millennium Prize problem has actually fallen, which is why the Clay Mathematics Institute’s “active” status still stands.
The October 7 Correction: Three Manuscripts Withdrawn
That gap between automated verification and full acceptance showed up almost immediately. OpenAI confirmed on October 7 that three of the original 722 manuscripts had to be pulled entirely: a sign error that invalidated a stabilization-trace argument killed that result outright, and two further manuscripts whose constructions depended on it were withdrawn along with it. A separate batch of 14 manuscripts was revised rather than withdrawn, with 13 citations updated, correcting smaller issues that didn’t invalidate the underlying results. Taken together, the catalog’s standing count fell from 722 to 719 papers within 24 hours of the original release. For a company leaning on automated proof-checking as evidence of rigor, a sign error slipping through and cascading into two dependent papers is close to the exact failure mode outside mathematicians warned about: a single mistake that propagates through everything built on top of it before anyone catches it. That OpenAI caught and disclosed the error within a day is arguably the more reassuring part of the story, since it shows the verification pipeline catching real problems rather than only rubber-stamping convenient results.
Lean Formalization Reaches 42% of the Catalog
The formal-verification side of the ledger moved in the other direction. AI Weekly reported on October 13, in coverage relayed by The NovTech, that six new Lean formalizations had been added to the catalog since the October 6 release. Lean is the same style of machine-checkable proof-verification language referenced above, the kind of tool that can confirm a logical argument follows valid inference rules without a human re-deriving every step by hand. Those six additions pushed the share of the catalog with a completed Lean formalization to roughly 42%, or about 300 of the 719 manuscripts still standing after the October 7 withdrawals, a count that matches OpenAI’s own October 7 revision log and that Terence Tao has separately referenced. That leaves well over half the catalog, around 419 manuscripts, without a formal machine-checked proof behind it yet, relying instead on the automated verification method OpenAI described to Nature. The 42% figure is a more concrete yardstick for the field’s actual confidence level than the raw manuscript count ever was: it measures how much of the October release has cleared a second, independent layer of checking, rather than how much OpenAI chose to publish.
A Timeline: From 10 Problems in August to 719 Manuscripts in October
The August-to-October Escalation
The scale of OpenAI’s math claims escalated fast over a roughly ten-week stretch. American Bazaar Online reported that an August 2026 disclosure involved an internal model producing results tied to 10 major mathematical problems, including a construction establishing the existence of non-sofic groups in group theory, work touching Erdős problems 146 and 180 in extremal graph theory, and a superexponential lower bound for multicolor triangle Ramsey numbers tied to Erdős problem 183.
Three Developments in the Same Week
Business Standard then reported that training for the model behind the broader claim began on August 28, and that within less than a month OpenAI said the system had solved or made substantial progress on more than 100 longstanding problems, a claim that directly prompted the September formation of the Tao-linked advisory group. By October 6, the scope had widened again to the full 722-manuscript, 372-result-family release covering roughly 4,000 tested problems, as reported by GIGAZINE and The Indian Express. Three developments followed in quick succession: a sign error forced the withdrawal of three manuscripts on October 7, OpenAI’s own October 7 revision log put Lean-formalized coverage at roughly 42% of the surviving catalog, and the Advisory Group on Mathematics and AI publicly criticized the release’s transparency between October 8 and October 16. All three are detailed in the table below.
| Date (2026) | Claim | Reported Scale | Verification Status | Source |
|---|---|---|---|---|
| August | 10 major problems, incl. non-sofic groups, Erdős 146/180/183 | 10 problems | Not independently confirmed | American Bazaar Online |
| Late August | Model training begins | N/A | N/A | Business Standard |
| September | 100+ problems solved or advanced since Aug. 28 training start | 100+ problems | Prompted advisory group formation | Business Standard |
| September | Navier-Stokes existence/smoothness claim via ~10,000 AI agents, ~88 hours | 1 Millennium Prize problem | Clay Mathematics Institute lists it “active,” not solved | BBC, CNN, Galaxy Research |
| September 21 | Advisory Group on Mathematics and AI formed, linked to IAS | N/A | Coordinating review process | Terence Tao’s blog |
| October 6 | Full manuscript release: 722 manuscripts, 372 result families | ~4,000 problems tested | Automated verification claimed; independent review ongoing | OpenAI, GIGAZINE, The Indian Express |
| October 7 | Sign error invalidates 1 proof + 2 dependent papers; 3 manuscripts withdrawn, 14 revised | 719 manuscripts remain | Withdrawal confirmed by OpenAI | OpenAI |
| October 13 | 6 new Lean formalizations added to the catalog | ~42% of catalog formalized (300 of 719) | Formal verification progress reported | AI Weekly, via The NovTech |
| October 8-16 | Advisory group criticizes release transparency; requests machine-readable links between proofs and formal artifacts | 10 of 719 manuscripts (~1.4%) include chain-of-thought material | OpenAI had not supplied the requested metadata as of this reporting | TechCrunch, AGMAI |
What Mathematicians Are Actually Worried About
Strip away the headline numbers and the mathematics community’s objection comes down to process, not pride. Mathematics has operated for centuries on a slow, adversarial review cycle: a proof circulates, specialists try to break it, gaps get patched or the claim collapses, and only after that gauntlet does a result become accepted. OpenAI’s release model, publish broadly and let the advisory group help coordinate how it’s framed, inverts the usual order. The manuscripts are public before the community-wide verification process has run its course.
The Verification Backlog Problem
That is precisely the “negative externality” flagged in the open letter OpenAI itself cited. If AI labs compete on who can claim the most solved open problems, the incentive to publish fast can outpace the incentive to get every result exactly right, and the burden of catching an error shifts onto an already time-strapped volunteer community of specialists. Tao’s advisory group appears to be an attempt to manage that risk by giving OpenAI’s release pipeline mathematician input before manuscripts hit GitHub, rather than only after the fact.
The October 7 withdrawal is a small-scale preview of exactly this risk: a sign error that invalidated one proof and cascaded into two more before anyone outside OpenAI had a chance to review the full catalog line by line. Catching it within 24 hours is a good outcome, but it happened because OpenAI was still actively re-checking its own release. The entire point of the open letter’s complaint is that an outside volunteer community, nine unpaid mathematicians among them, has far less bandwidth to catch the next one.
Tao’s Advisory Group Flags a Chain-of-Thought Gap
The clearest evidence that this isn’t a hypothetical risk arrived in mid-October, when the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), the same Tao-linked panel OpenAI formed in September, turned its coordination role into public criticism. Between October 8 and October 16, AGMAI said that only 10 of the 719 standing manuscripts, roughly 1.4%, included the chain-of-thought material showing how the model actually arrived at each result, according to TechCrunch’s reporting on the group’s statement. That’s a different gap than the Lean-formalization numbers covered above. A completed Lean formalization confirms the final proof’s logical steps check out. Chain-of-thought material shows the path the model took to get there: the intermediate reasoning, dead ends, and corrections behind the final answer. OpenAI’s catalog has plenty of the former and almost none of the latter.
Why that distinction matters to working mathematicians: a reviewer checking whether a 719-manuscript catalog is trustworthy at scale needs more than a pass-or-fail formal check on each final proof. Without the reasoning trace, there’s no way to tell whether the model solved a problem through sound mathematical reasoning that happens to generalize, or arrived at a valid answer through some other path a formal checker can’t distinguish from genuine insight. Tao’s group raising this gap in public, rather than negotiating it privately before the October 6 release went out, is itself notable. It suggests the advisory group’s “coordination” mandate, as Tao described it back in September, carries enough independence to push back publicly on the lab it’s supposed to be helping.
The Machine-Readable Metadata OpenAI Hasn’t Provided
AGMAI’s criticism came with a specific, actionable request attached, not just a general complaint about opacity. The group asked OpenAI for machine-readable links connecting each natural-language proof to its corresponding formal verification artifact, the kind of structured metadata that would let reviewers, and automated tooling, check which of the 719 manuscripts have cleared Lean formalization without manually cross-referencing a GitHub repository line by line. As of AGMAI’s public statement, OpenAI had not provided that metadata, according to both TechCrunch’s coverage and the advisory group’s own account.
The request sits right next to, but distinct from, the 42% Lean-coverage figure discussed earlier in this piece. AGMAI isn’t disputing that roughly 300 of the 719 manuscripts have a completed formalization. It’s pointing out that even where that work exists, there’s no structured way to confirm which specific manuscripts those are without the kind of manual digging that let the October 7 sign error sit unnoticed for as long as it did. A verification pipeline that can’t be queried programmatically scales worse than the manuscript count it’s supposed to be checking, and that mismatch, publishing at machine speed while verifying at human speed, is the same tension running through every section of this story.
What the Verified Record Says About the 719 Manuscripts
The reported collection, published on October 6, 2026, included claimed progress on three Millennium Prize problems and a claimed proof of the Unique Games Conjecture. The manuscripts claimed full or partial solutions to more than 370 open mathematical problems, according to available reporting, so saying that OpenAI solved 370 open problems is misleading, because verification was incomplete. It is equally inaccurate to describe the collection as 719 verified mathematical breakthroughs.
Reporting attributes to Tao the view that one result concerning a zero-free region for Dirichlet L-functions represented the field’s largest advance on the Riemann Hypothesis in about 50 years, while not proving the Riemann Hypothesis. The reported 7/8 zero-free-region result should therefore not be read as a proof of the Riemann Hypothesis, and the claim that OpenAI solved it is false on the available evidence. Tao also should not be described as having verified all 719 manuscripts.
About 300 of the 719 manuscripts were reported as having Lean-related formalization materials, but that does not establish that all underlying claims were independently understood, peer-reviewed, or correct. The release remains publicly discussed, yet the correctness of all 719 manuscripts has not been established. For readers, the practical approach is to treat each manuscript as a claim awaiting review and to check which results have independent confirmation before citing them.
Market and Industry Impact
The Credibility Stakes for OpenAI
For OpenAI, mathematics has become a proxy battleground for the much bigger argument about whether frontier models are approaching genuine reasoning capability or just producing plausible-looking text at scale. A verified, field-accepted resolution of even one Millennium Prize problem would be a watershed moment for the company’s research credibility, arguably bigger than any product launch. That’s also exactly why the stakes of getting verification wrong are so high: a high-profile claim that later unravels under scrutiny would hand ammunition to every critic arguing that AI math “breakthroughs” are overstated.
Ripple Effects Across the AI Industry
The broader AI industry is watching closely because math-benchmark bragging rights have become a competitive currency. Labs increasingly cite olympiad-style and open-problem performance the way they once cited standard coding and reasoning benchmarks. If OpenAI’s advisory-group model, publish with outside mathematician coordination rather than purely in-house validation, becomes the norm, it could slow down the pace of splashy math announcements industry-wide in favor of more defensible, slower-vetted claims. That would be a meaningful shift after a year in which AI labs have raced to attach their names to famous unsolved problems.
Spillover Into the Enterprise Business
There’s also a credibility spillover risk for OpenAI’s enterprise and developer business. Readers who’ve followed the company’s GPT-6.1 Sol API rollout or its broader model lineup know OpenAI markets reasoning capability as a core selling point against rivals. If the math claims face a credibility hit, that narrative takes damage well beyond the mathematics department.
Competitive Landscape: Who Else Is Racing Into AI Mathematics
OpenAI is not alone in chasing mathematical prestige with frontier models. Google DeepMind’s AlphaProof and AlphaGeometry systems have previously reached medal-level performance on International Mathematical Olympiad problems, work that is tracked and judged against the IMO’s own published problem sets and standards, as maintained by the International Mathematical Olympiad’s official site. That olympiad context matters as a comparison point: IMO problems have fixed, publicly known correct answers and a formal judging panel, which is a far more constrained and checkable environment than an open research problem with no known answer key.
Why Olympiad Problems Are Easier to Judge
The distinction is the heart of why OpenAI’s current situation is harder to referee than an olympiad scorecard. Solving a competition problem with a known answer is an achievement, but it is fundamentally different from generating a novel proof for a problem nobody has ever solved, where there’s no answer key to check against, only the slow process of expert review. Readers tracking the wider model race, including coverage of GPT-6 Astra against Gemini 4 Argon and DeepSeek, will recognize the pattern: benchmark claims move fast, and the harder, slower work of independent confirmation lags behind.
| Approach | Example | Problem Type | Verification Method | Independently Confirmed? |
|---|---|---|---|---|
| OpenAI internal model, Oct. 2026 release | 719 manuscripts, 372 result families | Open research problems, no known answer | Automated proof-checking tool, per OpenAI (via Nature) | Under review by outside mathematicians |
| OpenAI Navier-Stokes claim | 10,000-agent, 88-hour effort | Millennium Prize problem | Automated verification claimed | No; Clay Mathematics Institute lists it “active” |
| Google DeepMind AlphaProof/AlphaGeometry | IMO-style problem solving | Competition problems with known answers | Formal IMO judging panel criteria | Yes, for medal-level competition scoring |
| Traditional peer review | Standard academic mathematics | Open research problems | Specialist referees, journal review cycle | Yes, but typically over months to years |
Historical Context: AI’s Long Run at Mathematics
AI’s interest in mathematics is not new, but the scale of OpenAI’s 2026 push is a departure from the field’s earlier, narrower wins. For years, machine learning’s mathematical headlines centered on olympiad-style benchmarks, systems trained to match or beat top human competitors on problems with known, checkable answers. That framing was useful precisely because it sidestepped the verification problem: judges already knew what a correct answer looked like.
From Benchmark Wins to Open-Problem Claims
What changed across 2026 is the jump from “solve a known problem faster” to “generate novel results on problems nobody has solved.” That’s a fundamentally harder claim to substantiate, because there’s no answer key, only the painstaking process of expert mathematicians working through the logic line by line, or, increasingly, running it through automated proof-checking software. OpenAI’s August-to-October run, from 10 problems to more than 100 to a 722-manuscript release trimmed to 719 within a day, tracks an industry-wide pattern of AI labs moving their ambitions from “beat the benchmark” to “extend the frontier,” and running into the field’s centuries-old insistence on independent proof before anyone gets to use the word “solved.”
What This Means for Developers and Researchers
For working mathematicians and computer scientists, the practical takeaway isn’t to dismiss the manuscripts outright. The Indian Express’s reporting that the model was tested against roughly 4,000 problems, with the average result requiring resources equivalent to about three hours of ChatGPT Pro reasoning, suggests OpenAI itself is treating this as a mass-generation-then-curation pipeline rather than a single hand-polished proof. Some fraction of 372 result families touching hundreds of open problems is likely to hold up; the open question is which fraction, and how fast the field can sort the sound results from the flawed ones.
Practical Guidance for Builders
That sorting process is exactly what Tao’s advisory group was formed to help manage, and it’s a template other AI labs will likely watch closely. Software engineers building on top of frontier reasoning models, including those following Claude Opus 5.5’s API rollout or weighing model choices covered in LMArena versus Artificial Analysis leaderboard comparisons, should treat headline math claims the way they’d treat any unverified benchmark score: directionally interesting, not yet load-bearing.
Predictions: Where the OpenAI Math Verification Story Goes Next
Verification and Transparency
- Expect the Clay Mathematics Institute’s “active” status on Navier-Stokes to persist for months, not weeks; Millennium Prize verification has historically taken years, and an AI-generated proof is unlikely to get an expedited track given the stakes of a $1 million prize and permanent mathematical record.
- The Advisory Group on Mathematics and AI will likely formalize a recurring review cadence rather than a one-time consultation, given Tao’s framing of the group’s role as ongoing “coordination” of releases rather than a single sign-off.
- AGMAI’s October request for machine-readable links between proofs and formal artifacts is likely to become a standing condition attached to future releases rather than a one-time complaint, since the group raised it publicly instead of resolving it before the catalog went live.
Industry and Competitive Response
- Rival labs, including Google DeepMind, will likely lean harder into checkable, answer-key benchmarks like IMO-style problems in the near term, since those claims are far easier to defend publicly than open-problem claims facing the same scrutiny OpenAI is now absorbing.
- More open letters from mathematicians are likely if additional labs announce open-problem “solutions” without a parallel outside-review mechanism in place, especially given the precedent the “Severe Misalignment of AI in Mathematics” letter has now set.
- Expect OpenAI to lean more heavily on automated proof-checking tools and Lean-style formalization in future releases, since reaching 42% formalized coverage within a day of the original release suggests the company is treating formal verification as the faster path to credibility than waiting on manual peer review alone.
The Bigger Picture for AI Credibility
What’s playing out between OpenAI and the mathematics community is a preview of a tension that’s going to recur across every scientific field frontier AI touches. The moment an AI lab claims a result in a domain with rigorous, slow, human-run verification standards, whether that’s mathematics, physics, or medicine, the lab inherits that domain’s verification timeline whether it wants to or not. OpenAI can publish 722 manuscripts in a day. It cannot force the mathematics community to certify them in a day, and Tao’s advisory group exists precisely because that mismatch needed a formal structure before, not after, it became a crisis.
Why Verification Can’t Be Rushed
The company’s October 6 release is best read as a milestone in volume, not yet a milestone in confirmed mathematical truth. Whether any individual result among those 372 families eventually earns the kind of unambiguous acceptance the field reserves for genuinely solved problems will play out over the coming months, likely one peer-reviewed confirmation or one public retraction at a time.
Frequently Asked Questions
Has OpenAI actually solved the Navier-Stokes problem?
Not according to the Clay Mathematics Institute, which, per Galaxy Research’s reporting, still lists the Navier-Stokes existence and smoothness problem as “active” rather than “solved.” OpenAI has proposed a result, but independent verification by outside mathematicians was still underway at the time of the most recent reporting.
What is the Advisory Group on Mathematics and Artificial Intelligence?
It’s a panel formed in September 2026, associated with the Institute for Advanced Study, that includes mathematician Terence Tao. According to Tao’s own description, the group advises OpenAI on how to coordinate the release of mathematical results the company reports were produced by its internal model.
How many manuscripts did OpenAI release in October 2026?
OpenAI’s October 6 release included 722 manuscripts organized into 372 result families, according to coverage from GIGAZINE and The Indian Express. That total fell to 719 on October 7, after a sign error forced the withdrawal of three manuscripts and prompted revisions to 14 others.
What is the “Severe Misalignment of AI in Mathematics” letter?
It’s an open letter from mathematicians, referenced on OpenAI’s own advisory-group page, raising concerns about the negative externalities of using open-problem-solving as a competitive benchmark for AI systems, including the risk that speed-driven publication outpaces the field’s ability to verify results.
How is OpenAI verifying its AI-generated math results?
According to Nature’s reporting, OpenAI says it used an automated verification method it describes as becoming the standard for rigor in mathematics, though this is distinct from the full independent peer-review process the broader mathematics community still expects before treating a result as formally confirmed.
How does this compare to Google DeepMind’s AI math work?
Google DeepMind’s AlphaProof and AlphaGeometry systems have focused on IMO-style competition problems with known, checkable answers judged against official International Mathematical Olympiad standards. That’s a more easily verified category of claim than OpenAI’s open-problem work, where no answer key exists and verification depends entirely on expert review.
When did OpenAI’s internal model start training for these results?
Business Standard reported that training began on August 28, 2026, and that within less than a month OpenAI said the model had solved or made substantial progress on more than 100 longstanding mathematical problems.
How many manuscripts remain in OpenAI’s catalog after the October withdrawals?
719. OpenAI’s October 6 release originally totaled 722 manuscripts, but a sign error invalidated one proof and two manuscripts that depended on it, and all three were withdrawn on October 7. A further 14 manuscripts were revised rather than pulled. Separately, six new Lean formalizations reported on October 13 brought the share of the catalog with a completed formal proof to roughly 42%, or about 300 of the 719 standing manuscripts.
Who are the members of OpenAI’s Advisory Group on Mathematics and Artificial Intelligence?
The group has 9 unpaid mathematicians in total. Terence Tao is the member who has spoken most publicly about its mandate, and the roster also includes Timothy Gowers, Martin Hairer, and Edward Witten, according to OpenAI’s own description of the panel. It was established in September 2026 to help coordinate how OpenAI releases and frames its mathematical results.
What did the Advisory Group on Mathematics and AI criticize about the October release?
Between October 8 and October 16, 2026, the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) publicly criticized the release’s transparency, according to TechCrunch and the group’s own statement. It noted that only 10 of the 719 manuscripts, about 1.4%, included chain-of-thought material showing how the model reached its results, and it asked OpenAI for machine-readable links tying each natural-language proof to its formal verification artifact. OpenAI had not supplied that metadata as of the group’s statement.
