OpenAI has published 372 claimed results in mathematics and theoretical computer science from an unreleased internal model, posting them on 6 October 2026 in a public GitHub repository alongside a company blog post [2][3]. The catalogue groups more than 700 manuscripts into result families, and OpenAI says many of the proofs come with Lean formalisations that a computer can check [2][3].
The release landed a month after a high-profile Navier–Stokes episode, and it tests how mathematicians can absorb hundreds of machine-made claims at once. Formal verification covers some of the proofs. Independent scrutiny of how they were produced has barely begun.
On 6 October 2026, Scientific American reported that OpenAI released 372 results from the same unreleased internal model that recently produced a major mathematics breakthrough, with each result said to resolve or make substantial progress on a major open question in mathematics or theoretical computer science. [1]
OpenAI says each result took about three hours of compute
According to Scientific American’s Joseph Howlett, OpenAI posted the results in a GitHub repository at 6 p.m. EDT. Claimed highlights mentioned in the article include a solution to the four-dimensional Kakeya conjecture, improvements on important computer algorithms, and progress toward the Riemann hypothesis. [1]
An OpenAI spokesperson told Scientific American that the new model—not released to the public—produced almost every result in response to a single prompt handed to a single AI agent, a contrast with the earlier Navier–Stokes work described as a 10,000-strong agentic swarm that cost millions of dollars in compute. The spokesperson also said some results might have taken multiple attempts. [1]
Those production-process claims are company statements. OpenAI’s own repository adds some detail: the model was posed approximately 4,000 problems over the evaluation, and on average each result used three hours of ChatGPT Pro thinking compute. It names exceptions to the fixed procedure, including work on a zero-free region for the Riemann zeta function, whose write-up was edited by humans for readability [3]. The blog says the release includes 10 summaries of the model’s reasoning, but not the prompts [2].
What Lean verification does—and does not—settle
Scientific American reports that many of the results have already been verified in Lean, a programming language that validates a proof’s logic, and are therefore “all but certain to be correct” as formal arguments. [1]
The repository is candid that coverage is partial. Its README says “not all have accompanying Lean formalizations” and that “some of the unformalized results could have issues”. A day after release, on 7 October, OpenAI’s change log recorded that a sign error in one paper invalidated an argument used by two others, and all three manuscripts were withdrawn. Fourteen more were revised with proof repairs or corrected statements. That brought the share of top-line results formalised to 300 of 719, about 42% [3].
Lean can check that a formalised proof follows from stated axioms and lemmas. It does not automatically settle novelty, importance, whether the formal statement matches the informal open problem, or whether OpenAI’s “single agent / single prompt” story is accurate.
Independent check pending
MIT mathematician Andrew Sutherland told Scientific American: until and unless OpenAI releases the model and people can replicate the results, claims about one-shotting problems with a single agent should be treated as unverified—“We should ask for receipts.” [1]
On 21 September 2026, Scientific American notes, OpenAI said it was assembling an independent advisory group of mathematicians on responsibly publishing AI-generated results. That group’s recommendations, as summarised in the article, call for making public the model, exact prompt and compute time behind each result. OpenAI is instead revealing average compute time for a problem plus some additional statistics—and no prompts. A spokesperson said the team is taking the guidelines seriously and doing its best to comply, but that the company is not bound by the recommendations. OpenAI also said it is working to release the model as quickly and responsibly as possible. [1]
Scientific American further reports that many of the newly released results are not yet understood by OpenAI’s own mathematicians, per the spokesperson. Mathematicians quoted in the piece disagree on pace and disclosure: Toronto’s Daniel Litt sees little reason to keep answers secret; Terence Tao has criticised frontier labs for an “insane” pace of AI-generated results. [1]
Questions the release leaves open
- Independent check pending: model, prompts and per-result compute are not fully public as the advisory recommendations urged.
- Lean-checked proofs are not the same as community acceptance of mathematical importance or originality.
- “Substantial progress” on open problems is a company framing until specialists finish reading.
- Cheaper-than-swarm generation, if true, would change access—but that “if” is exactly what Sutherland flags as unverified.
A flood of candidate proofs shifts the work onto human referees
The release changes the bottleneck in mathematics more than the frontier itself. Producing a plausible proof of a hard conjecture used to be the scarce part; checking it, understanding it and deciding whether it matters was work spread across a small community over years. OpenAI has inverted that ratio, dropping hundreds of manuscripts at once on referees who did not ask for them. Lean helps, because a formal proof that compiles cannot hide a logical gap, but the withdrawal of three papers within a day shows why the unformalised majority needs caution.
OpenAI deserves credit for publishing a change log, keeping withdrawn versions visible and deferring to mathematicians on credit and precedence. It has also moved to pay for the checking: on 15 October it added a $5m contribution to a Simons Foundation initiative to help mathematicians examine AI-produced results [2], part of the wider $15m maths AI commitment with Simons and Anthropic. That is the right direction, but it is small next to the volume of output, and the advisory group’s request for prompts and per-result compute remains unmet.
The contrast with Google DeepMind is instructive. AlphaProof Nexus went through peer review in Science with narrower, counted claims, while Axiom has leaned on full Lean formalisation as its evidence. OpenAI is betting on volume first and verification later. On 26 October it added a manuscript claiming to refute the Exponential Time Hypothesis, now being examined by an independent working group of computer scientists, which TSN covered in OpenAI says its AI refuted the Exponential Time Hypothesis [3]. What specialists conclude about that paper, and about the 58% of results with no formal proof, will decide how much of this catalogue survives.
Updated 29 October 2026: added OpenAI’s blog post and repository as primary sources, the 7 October withdrawals and 42% formalisation figure, the 15 October Simons funding and the 26 October additions.
Sources
- Scientific American (Joseph Howlett; edited by Lee Billings), “OpenAI unleashes hundreds more math results upon a field already in shock,” 6 October 2026 — https://www.scientificamerican.com/article/openai-unleashes-hundreds-more-math-results-upon-a-field-already-in-shock/
- OpenAI, “Sharing AI progress in mathematics,” 6 October 2026 (updated 15 October 2026) — https://openai.com/index/sharing-ai-progress-in-mathematics/
- OpenAI, openai/math GitHub repository: README and history (change log), accessed 29 October 2026 — https://github.com/openai/math
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterFollow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
