OpenAI dropped a research bomb on the mathematics world this week, and the fallout is still spreading. On October 6, 2026, the company published a sprawling GitHub repository containing 722 manuscripts generated by an unreleased internal model, organized into 372 result families and covering 377 selected mathematical results. By October 7, The Washington Post was reporting that mathematicians were “trying to make sense of OpenAI’s Tuesday release of findings on more than 300 open problems made by an internal artificial intelligence model.” The headline framing stuck: progress on more than 300 unsolved math research problems, in one bulk drop, with no public access to the model that produced it.
What makes this moment different from OpenAI’s usual cadence of benchmark announcements is the sheer scale and the lack of independent review before publication. The company did not wait for peer review, did not identify the model by name, and did not limit the claims to tidy competition problems. Instead it released a pile of papers spanning algebra, number theory, geometry, mathematical logic, topology, theoretical computer science and differential equations, then effectively handed the grading job to the global mathematics community. The reaction, documented across outlets from Seoul to Boston to Mumbai, mixes genuine awe with a specific and pointed discomfort about how science is supposed to work.
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 Released on October 6
The centerpiece of the announcement is a repository hosted at github.com/openai/math, which the company describes as containing mathematical manuscripts and supporting proof artifacts produced by an internal OpenAI model, the same 722-manuscript corpus detailed in earlier coverage of the secret model behind it. OpenAI has not disclosed the model’s name or made it available for outside testing, and it should not be assumed to be GPT-6 Sol, Luna, or Astra. According to The Verge, the release covers 722 manuscripts organized into 372 result families touching 377 distinct mathematical results. The repository also carries Lean formalizations, though OpenAI said the results were at different verification stages and that many, but not all, manuscripts had been formalized.
The scope stretches well beyond the kind of competition mathematics that AI systems have chased for years. The Boston Globe reported that the manuscripts span algebra, number theory, theoretical computer science, mathematical logic and topology, a spread that touches both pure and applied branches of the field. Earlier reporting from WIRED indicated the underlying model had been in training since August 28, 2026, and that OpenAI had already claimed progress on more than 100 long-standing open problems before the October release expanded that figure several times over.
Mashable’s coverage framed the dump bluntly: an “info dump that will take hundreds of human hours to sift through.” That phrase captures the core tension of this story. OpenAI generated the material quickly. Verifying it will not be quick at all.
The Numbers Behind the “More Than 300” Headline
Part of the confusion swirling around this story comes from the fact that several different numbers are circulating, and they are not interchangeable. The Washington Post’s headline figure of “more than 300” is a rounded, reader-friendly summary. The more granular numbers tell a layered story about how the release was built and counted.
| Metric | Reported Figure | Source |
|---|---|---|
| Candidate open research questions evaluated | Roughly 4,000 | Business Model Analyst |
| Selected mathematical results released | 377 | The Verge / shattered.io |
| Result families (grouped manuscripts) | 372 | OpenAI GitHub repository |
| Total manuscripts published | 722 (some trackers list 719) | GitHub / The Verge |
| Public headline framing | “More than 300” problems | The Washington Post |
| Earlier OpenAI claim (pre-October) | “More than 100” long-standing open problems | WIRED |
| Outside framing of difficulty tier | 90 of roughly 500 hardest open problems | Startup Fortune |
The 377-versus-372 gap that confused early readers is not necessarily a contradiction. The 377 figure refers to individually selected results or problem-level claims, while 372 refers to how OpenAI grouped related manuscripts into families, according to reporting from shattered.io. The deeper and more important number is the 4,000-problem pool OpenAI reportedly tested against after its model had already saturated existing math benchmarks. That detail matters because it signals OpenAI ran out of standard tests to measure the model against and had to build a new, harder evaluation set from genuinely unsolved questions.
The Model Nobody Outside OpenAI Has Touched
A detail that keeps surfacing in coverage of this release, and one that frustrates mathematicians more than the numbers themselves, is that OpenAI has not released, named, or opened the model to outside testing. Fortune reported that “OpenAI said it achieved the results using an unreleased internal AI model,” a framing echoed across nearly every outlet covering the story. Independent researchers cannot run their own prompts against the system, cannot probe its failure modes, and cannot check whether the 722 manuscripts represent a representative sample of the model’s output or a curated best-case selection.
That opacity is a familiar pattern for OpenAI, which has historically kept its most capable research systems internal while publishing selected outputs. It is also the single biggest reason the mathematics community’s reaction has tilted toward caution rather than celebration. A manuscript is not a verified theorem. It is a claim that requires line-by-line checking, and until that checking happens at scale, “progress on 300 problems” remains a company’s characterization of its own work rather than an independently confirmed scientific result.
Quasi-Riemann and Navier-Stokes: The Headline Claims
Among the hundreds of results, two stand out because they touch problems with decades of mathematical weight behind them. Reporting attributed to Chosun Biz’s coverage of Wall Street Journal reporting states that “the announcement included a proof of the quasi-Riemann hypothesis and progress related to three of the five Millennium Problems that remain unsolved.” The Millennium Prize Problems, set by the Clay Mathematics Institute in 2000, are widely regarded as the seven hardest open questions in mathematics; only one, the Poincaré conjecture, has been solved to date.
Separately, WIRED’s October 6 reporting referenced an earlier statement describing the model as having resolved the Navier-Stokes existence and smoothness problem, one of those same seven Millennium Problems, in addition to more than 100 other long-standing open questions. That statement dates to September 8, 2026, when OpenAI said an internal model had found finite-time blow-up solutions for the Navier-Stokes problem. It is worth stressing what “resolved” means in this context: a model-generated manuscript claiming a proof, not a community-accepted, independently verified solution. The distinction between those two things is exactly what has triggered the loudest pushback from working mathematicians.
Why Mathematicians Are Stunned — and Why They’re Worried
The Washington Post’s framing, that the release is “stunning humans,” captures one half of the reaction. The other half is anxiety about what happens to the practice of mathematics if a company can generate hundreds of plausible-looking proofs faster than the global community of professional mathematicians can check even a fraction of them. The News (Pakistan) reported that OpenAI is framing this as a “new era of discovery,” a phrase that sits uneasily with a field built on painstaking, line-by-line verification.
South Korea’s SBS News, citing Wall Street Journal reporting, described the scale of the shock directly: “If a human did this, Fields Medal instantly.” That line, repeated across multiple regional outlets covering the U.S. story, gets at the unsettled feeling running through the field. A human mathematician who solved even a handful of century-old open problems in one sitting would be a historic figure. A model that claims to have made progress on hundreds at once, without independent confirmation, forces the field to ask what the achievement actually means before it can celebrate it.
The Fields Medalists Pushing Back
The skepticism is not limited to anonymous online commentary. According to SBS News’ citation of Wall Street Journal reporting, a prior controversy around OpenAI’s math claims involved 25 Fields Medalists, who signed an open letter warning that AI’s growing role in mathematical problem-solving could undermine the field’s standards and sense of purpose. The Fields Medal is mathematics’ highest honor, awarded every four years to a small handful of researchers under 40, which makes a 25-signature letter from that group an unusually concentrated show of institutional concern.
That letter predates the October 6 release but frames exactly why this week’s dump landed the way it did. Reports say the 25 Fields Medalists signed it on September 11, 2026, under the title “A Severe Misalignment of AI in Mathematics,” and on October 7, 2026 the Association for Human Mathematics went further, describing the release as “a demonstration of power” rather than scholarship and urging mathematicians to stop working with OpenAI. The worry is not that AI-generated mathematics is worthless. It is that a flood of unverified claims, released faster than the community can check them, risks eroding the trust and incentive structure that peer review is built to protect. If manuscripts pile up faster than verification capacity, the field either has to slow down the AI labs, scale up verification dramatically, or accept a growing backlog of unconfirmed claims sitting in public repositories.
AGMAI: OpenAI’s Answer to the Trust Problem
OpenAI’s response to that exact criticism appears to be the creation of an independent advisory group referred to in coverage as AGMAI, the Advisory Group on Mathematics and Artificial Intelligence. The Verge reported that AGMAI, described as a newly formed independent group of elite mathematicians, characterized the release as involving solutions to “hundreds” of open questions while emphasizing responsible communication around the claims. WIRED separately noted that the Institute for Advanced Study’s advisory circle was consulted on how OpenAI should release and communicate results of this scale.
The existence of AGMAI suggests OpenAI anticipated the backlash and tried to get ahead of it by building in outside mathematical review before publication, rather than after. Whether that group reviewed all 722 manuscripts or only a representative sample is not clear from available reporting, and that ambiguity is itself part of what critics are pressing on.
A Timeline: From One Prompt to 722 Manuscripts
This release did not appear out of nowhere. It is the latest and largest escalation in a string of OpenAI mathematics claims stretching back to late August. Readers following the company’s earlier 377-versus-372 results dispute and the mathematician scrutiny of the 722 claims will recognize the pattern: a big number lands, then the counting methodology gets picked apart in public.
| Date | Event |
|---|---|
| August 28, 2026 | OpenAI begins training the internal model later tied to its math claims, per WIRED |
| September 2026 | OpenAI says the model resolved more than 100 long-standing open problems, including Navier-Stokes |
| October 6, 2026 | OpenAI publishes the 722-manuscript, 372-family repository on GitHub |
| October 7, 2026 | The Washington Post, Fortune and others report the “more than 300 problems” framing; OpenAI publishes a blog post titled “Sharing AI progress in mathematics” |
| October 8, 2026 | Follow-up coverage details the 377/372 counting dispute and the scale of the verification challenge facing mathematicians |
Compared with the earlier headline claim of “more than 100” problems, the October release is a different order of magnitude in raw document volume, even if the underlying mathematical substance has not yet been independently validated at the same scale. Some outside trackers, cited by Startup Fortune, framed the release as touching roughly 90 of the 500 hardest open problems in mathematics, a characterization OpenAI itself has not confirmed in those exact terms.
What Independent Verification Actually Requires
It helps to be concrete about why “verification” is not a quick formality. Checking a single nontrivial mathematical proof can take an expert reviewer days or weeks, especially when the proof introduces unfamiliar lemmas, leans on computational steps, or spans multiple sub-results the way many of OpenAI’s manuscripts reportedly do. Multiply that by 722 documents across seventeen reported subfields of mathematics, and the review burden stretches into the thousands of expert-hours, distributed across a global community of researchers who have their own existing work to do. The cost asymmetry is stark: Klu.ai reported that running OpenAI’s newer GPT-6.1 Sol model against the ArXivMath benchmark in September 2026 cost just $0.93 per problem, meaning machine-generated attempts at a problem are essentially free to produce at scale, while the human side of checking each one still runs on expert-days, not cents.
That is the gap The Boston Globe pointed to when it described the release as “further roiling” the field rather than settling it. Until independent mathematicians work through the backlog, every headline number in this story, 377, 372, 722, “more than 300,” describes the size of the claim, not the size of the confirmed result. OpenAI itself has reportedly acknowledged that outside experts will likely find errors and request revisions, building a correction process into the release rather than presenting it as a finished product.
Where Verification of the October 6 Release Stands
OpenAI published more than 700 mathematical files on October 6, 2026, but said the results were at different verification stages and that many, but not all, manuscripts had been formalized. Reports describe 722 manuscripts organized into 372 result families, with Lean formalizations, in a repository identified as openai/math. The available evidence describes manuscripts and partial formalization, not completed peer review, so the release should not be described as 722 independently peer-reviewed papers.
The model behind the work is still unidentified. The identity of the unreleased model used for the October 6 mathematics work was not disclosed, and it should not be assumed to be GPT-6 Sol, Luna, or Astra. On the human side, the Association for Human Mathematics described the release on October 7, 2026 as a “demonstration of power” rather than scholarship and urged mathematicians to stop working with OpenAI. No available source confirms that the medalists’ criticism has been withdrawn, resolved, or converted into a formal institutional partnership. For readers, the practical approach is to treat each manuscript as a claim until an independent mathematician confirms it, and to check whether a given result carries a Lean formalization before relying on it.
How This Compares to Google DeepMind and Anthropic’s Approach
OpenAI is not the only lab racing toward machine-generated mathematics, but its approach this week stands apart from its closest rivals in both scale and openness. Google DeepMind’s mathematics-focused systems have historically targeted narrower, formally verifiable problem sets, such as competition-style geometry and olympiad problems, where a proof can be mechanically checked against a formal system rather than read and judged by a human expert. That narrower scope trades breadth for built-in verification, the opposite tradeoff OpenAI appears to have made with its October release.
Anthropic, for its part, has leaned toward publishing specific, bounded results rather than bulk releases, a contrast visible in its own reported progress on a Riemann zeta function bound, which was presented as a single, discrete claim rather than a 722-document repository. Neither DeepMind nor Anthropic has, according to available reporting, attempted anything at the scale of OpenAI’s October drop, which makes the current controversy as much about release strategy as about the underlying mathematics. A smaller, slower, more verifiable release invites less skepticism; a larger, faster, less verifiable one invites more attention and more scrutiny in roughly equal measure.
Global Media Reaction: From Washington to Seoul to Karachi
Few AI stories this year have traveled across as many national media markets as this one. The Washington Post set the English-language framing that most outlets echoed. South Korea’s SBS News folded in Wall Street Journal reporting and the Fields Medalists detail for a domestic audience already following the broader AI-agent security stories out of the region’s banking sector. Pakistan’s The News emphasized OpenAI’s own “new era of discovery” framing. Indian outlet The Print tied the release to OpenAI’s earlier claim of cracking a 90-year-old math puzzle, presenting the October drop as the “mother lode” that followed. South Korea’s Chosun Biz, drawing on Wall Street Journal sourcing, highlighted the quasi-Riemann and Millennium Problem angle for a business-and-technology readership.
That spread matters beyond novelty. It shows this story landed simultaneously as a technology story, a science story, and in several markets, a business story, because OpenAI’s valuation and competitive position increasingly rest on claims about frontier model capability that extend well past chatbots and coding assistants into domains, like unsolved mathematics, that have no obvious commercial product attached yet.
Historical Context: AI’s Long Road Toward Real Mathematics
Machine assistance in mathematics is not new. Computer-assisted proofs date back decades, most famously the 1976 four-color theorem proof, which required a computer to check thousands of cases a human could not feasibly verify by hand. What is new with large language model-based systems is the shift from computers checking human-designed proofs to AI systems generating the proofs themselves, in natural mathematical language, across an open-ended range of subfields rather than one narrowly defined problem. OpenAI’s own line reflects that arc, running from o3-mini, released in January 2025 and optimized for science, math, and coding, to the unreleased research model behind this release.
That shift is also why this release reads differently than a typical AI capability announcement. Benchmark scores on coding or reasoning tests, the kind that accompany most model launches, are comparatively easy to verify: run the test, check the answer. Mathematical proofs at the research frontier do not work that way. There is no answer key, a point made explicitly in Business Model Analyst’s framing of the story: OpenAI’s model reportedly maxed out existing benchmarks (OpenAI itself reported 98% on FrontierMath Tier 4 with GPT-6 Astra in April 2026), so the company built a new test out of roughly 4,000 problems nobody has ever solved, and the resulting 722 papers now need human mathematicians to grade them by hand. Even the field’s dedicated evaluation efforts are tiny by comparison: MathArena’s own August 2026 leaderboard, published September 15, 2026, ran on just 57 questions, a scale that makes OpenAI’s 4,000-problem internal test set look less like a benchmark and more like an entirely new category of evaluation.
Market and Industry Impact
OpenAI is not a publicly traded company, so there is no direct stock price to track against this announcement. Available reporting does not establish a confirmed, specific market reaction among publicly traded AI-adjacent companies tied directly to this release. What the story does appear to be shifting is narrative and positioning rather than share price: it lands just weeks after OpenAI’s GPT-6 rollout, followed by the September 29, 2026 release of GPT-6.1 Sol according to Klu.ai, and amid broader industry jockeying over which lab can credibly claim the frontier of AI reasoning capability, a competition that increasingly uses hard scientific problems, rather than consumer chatbot features, as the proving ground.
That positioning fits a broader pattern of frontier labs making sweeping capability claims this year, echoed in OpenAI chief executive statements elsewhere about AI progress reaching a turning point. That positioning matters commercially even without a direct stock move. Enterprise and research customers evaluating frontier AI systems for scientific and engineering work are watching claims like this one as a signal of what the next generation of research tools might do, and how much independent verification those claims will require before they can be trusted in production research pipelines. The same dynamic the Fields Medalists flagged in their open letter, speed of generation outpacing capacity for verification, is exactly the risk enterprise buyers will need to budget for if they start relying on similar systems internally.
What Happens Next: Five Predictions
- A slow-motion verification scorecard emerges. Expect mathematicians to publicly confirm, dispute, or flag errors in specific manuscripts over the coming months rather than all at once, with outlets like shattered.io and The Verge likely to track the running tally.
- More open letters follow. Given the precedent of the 25-signature Fields Medalist letter, additional statements from mathematical societies or university departments are plausible if verification lags the pace of new claims.
- Rival labs respond with their own math claims. Google DeepMind and Anthropic have strong incentive to counter-position with their own frontier mathematics results, likely framed around narrower, more immediately verifiable claims to contrast with OpenAI’s bulk approach.
- AGMAI’s role expands or comes under scrutiny. The advisory group OpenAI stood up to vet these claims will likely face questions about its independence, its review process, and how much of the 722-manuscript set it actually reviewed before publication.
- Pressure builds for a formal verification standard. Expect renewed calls, echoing the broader AI-safety and AI-accountability debates already playing out in Washington, for standardized third-party review processes before AI-generated scientific claims are treated as established results.
The Bigger Picture for AI and Scientific Discovery
Strip away the specific numbers and this week’s story is really about a collision between two different speeds. AI labs can now generate plausible, technically sophisticated mathematical content faster than any point in history. The institutions built to check that content, peer review, academic publishing, expert consensus, move at the speed they always have, because checking rigorous proof is inherently slow, careful work that resists automation as completely as the generation step now embraces it.
OpenAI’s bet, visible in both the scale of the October 6 release and the creation of AGMAI, appears to be that the field will adapt its verification capacity rather than demand the company slow down generation. The 25 Fields Medalists who signed the earlier open letter are, in effect, betting the opposite: that without some brake on the pace of claims, the field’s core trust mechanism, independent verification before acceptance, risks being overwhelmed. Which bet wins will shape not just how mathematicians treat OpenAI’s next release, but how every scientific field responds when an AI lab shows up with hundreds of claims and no answer key.
Frequently Asked Questions
Did OpenAI actually solve 300 math problems?
OpenAI released manuscripts claiming progress on more than 300 open mathematical problems, according to The Washington Post. Available reporting does not establish that all of these claims have been independently verified as correct, complete solutions. “Progress” and “results” are the terms used in official OpenAI communication rather than a confirmed count of fully solved problems.
What model did OpenAI use to generate these results?
OpenAI has not publicly named the model. Multiple outlets, including Fortune and The Washington Post, describe it only as an unreleased internal AI model not available for outside testing.
Why is there a difference between the numbers 377, 372, and 722?
According to shattered.io and The Verge, 377 refers to selected mathematical results, 372 refers to the number of grouped result families, and 722 refers to the total number of manuscripts published across those families. They describe different layers of the same release, not conflicting claims.
Did OpenAI solve a Millennium Prize Problem?
Reporting attributed to Wall Street Journal sourcing via Chosun Biz describes a proof of the quasi-Riemann hypothesis and progress on three of the five still-unsolved Millennium Prize Problems. These are described as claims within the released manuscripts, which require independent mathematical verification before they can be treated as confirmed solutions.
Who are the Fields Medalists criticizing OpenAI?
SBS News, citing Wall Street Journal reporting, describes an open letter signed by 25 Fields Medalists warning that AI’s expanding role in mathematics could undermine the field’s standards. Available reporting does not provide a full list of individual signatories for this piece.
What is AGMAI?
AGMAI, the Advisory Group on Mathematics and Artificial Intelligence, is described by The Verge and WIRED as an independent group of mathematicians OpenAI consulted on how to release and communicate its mathematical results responsibly.
How long will it take mathematicians to verify these claims?
No official timeline has been reported. Given that checking a single complex proof can take an expert days or weeks, and the release spans 722 manuscripts across multiple subfields, independent verification is expected to unfold over months rather than days.
Is this related to OpenAI’s earlier claim about a 90-year-old math puzzle?
Yes. The Print’s coverage frames the October 6 release as following OpenAI’s earlier claim of cracking a decades-old mathematical puzzle, describing the new, larger release as building directly on that earlier announcement.
