On 6 October 2026 OpenAI published hundreds of mathematical results produced by a still unreleased internal model. According to an OpenAI spokesperson, almost all of them came from a single prompt to a single agent. Among the results is supposed to be a proof of a conjecture by Subhash Khot, the Unique Games Conjecture. Whether it is correct, nobody can yet say with certainty.
What happened
On 6 October 2026 OpenAI presented 372 so-called result families on GitHub. One report describes them as work from 17 areas of mathematics and theoretical computer science. The repository lists 719 manuscripts, whereby an initial count assumed 722 and three manuscripts were withdrawn after a correction. The 372 are thus result families and not individual problems, as Technology.org stresses.
One prompt, one agent
An OpenAI spokesperson told Scientific American that the model produced almost every result from a single prompt to a single agent; some results, however, needed several attempts. The Decoder puts the average computing effort at about three hours of ChatGPT Pro time per result.
The proof of the Unique Games Conjecture
Scott Aaronson writes on his blog Shtetl-Optimized that the release includes a proof of Subhash Khot's conjecture, the Unique Games Conjecture, and that it has a Lean certificate. His post calls the day a huge day for mathematics. He is not entirely sure, though: “Or at least, we’re pretty sure that it’s a proof!” He also notes that so far no human has understood just about any of these proofs; the race to do so has only just begun.
Aaronson's wife Dana Moshkovitz, who has worked on the conjecture for years, is reported by Technology.org to have found the submitted proof so hard to read that she could not read it without AI help.
Background
Lean is a proof assistant that checks the logical steps of a proof by computer. A Lean certificate is therefore a strong indication, but not understanding: one commentator points out that the check is reassuring but has nothing to do with understanding. Technology.org adds that Lean checks only the formal statement, not necessarily whether it matches the English text of the claim.
How many results have actually been formally checked is reported differently by the sources. Implicator gives about 42% of the main results, roughly 300 of 719, as having Lean formalisation, as of 7 October; that figure applies after the correction in which three manuscripts were withdrawn. AI Weekly gives only 162 of 722 manuscripts with proofs checked in Lean. So there is no single figure.
The release triggered sharp reactions. Andrew Sutherland of MIT said, according to AI Weekly, that claims of solving problems with a single agent should be treated as unverified until the model is released and the results can be reproduced. The same source reports that OpenAI disabled the “GitHub Issues” feature and withheld prompts that an advisory group had requested. The Decoder reports that some mathematicians accused OpenAI of ignoring the core of an advisory group's recommendation, which said that hard mathematical problems should not be tried on internal models, and called for a boycott.
There is also independent work on the same subject. According to one report, Dor Minzer's group at MIT posted a related “4-to-1” version of Khot's conjecture on ECCC in September, after rumours of an OpenAI proof had reached them.
What this means
For mathematicians, a problem of volume arises: when hundreds of manuscripts appear at once and hardly anyone understands them, reading and checking becomes the bottleneck. Aaronson's remark that the race to understand has only just begun describes exactly that.
For anyone who has to put AI news in context, the announcement is a lesson in caution. The claim “one prompt, one agent” comes from OpenAI; the formal check covers, depending on the count, only part of the results; and the model itself is not released, so nobody outside can trace how the results came about. Whether a proof counts as established is decided not by a certificate or a press release but by experts being able to follow it.
For research culture, the reactions show that it is not only about technology. The accusation that an advisory group's recommendation was ignored, and the call for a boycott, show that parts of the profession also criticise how the release was handled.
A note on interpretation: the information relies on news summaries and blogs. The sources evaluated found no independent check of the Unique Games proof itself, and the counts differ from source to source.
Frequently asked questions
What exactly does “372 results” mean?
It means 372 result families from 17 areas of mathematics and theoretical computer science, not 372 individual problems. The repository lists 719 manuscripts; an initial count said 722, and after a correction three manuscripts were withdrawn.
Has the Unique Games Conjecture been proved?
OpenAI presents a proof, according to Scott Aaronson with a Lean certificate. Aaronson himself writes, however, that he and others are “pretty sure” it is a proof; no human has understood the proofs so far. The sources evaluated found no independent confirmation.
What is Lean?
Lean is a proof assistant that checks the logical steps of a proof by computer. It checks only the formal statement, though, not necessarily whether it matches the English text of the claim.
Why is the announcement disputed?
Because it is unclear how many results have been formally checked, and because, according to one report, OpenAI withheld prompts that an advisory group had requested. A mathematician from MIT says the “one prompt, one agent” claim should only be believed once it can be reproduced.
