Claimed AI-led

719 manuscripts in 372 result families from an unreleased OpenAI model

OpenAI published 722 manuscripts, 719 after three were withdrawn the next day, grouped into 372 claimed results across number theory, algebraic geometry, analysis, combinatorics, theoretical computer science and mathematical physics, all produced by an unreleased internal model, with Lean formalizations initially attached to 235 of the families; this entry's 2026-10-07 assessment recorded no outside checks, and later external checker reports are described below.

Model
Unreleased internal OpenAI model, not named in the release
Field
Mathematics
Date
2026-10-06

What was found

The repository's README says the model was posed about 4,000 open problems during an evaluation, that each result used on average three hours of ChatGPT Pro thinking compute, and that the output was aggregated into families and filtered for significance; a family groups a principal result with companions, consequences or alternative proofs. The families claimed include the full Birch-Swinnerton-Dyer formula in Selmer corank at most one, the quasi-Riemann hypothesis, a negative answer to Hilbert's tenth problem over the rationals, the Unique Games Conjecture, 𝖫=𝖱𝖫=𝖡𝖯𝖫\mathsf{L}=\mathsf{RL}=\mathsf{BPL}, matrix multiplication with exponent at most 9/49/4, Kakeya in three and four dimensions, nonamenability of Thompson's group F, isomorphism of the free group factors and Kadison's similarity conjecture, alongside dozens of counterexamples to named conjectures. Lean 4 formalizations are linked from 235 families; the formalization catalogue lists 162 papers whose main result is formalized and Comparator challenge files for checking the statements against permitted axioms. Ten families also come with abridged summaries of the model's reasoning. The README names two exceptions to the fixed procedure, the work on a zero-free region for the Riemann zeta function and the Hodge conjecture for CM abelian varieties, and says the write-up of the ℜs>11/12\Re s > 11/12 zero-free region was edited by humans for readability. On 2026-10-07 OpenAI withdrew three manuscripts after a sign error in 'Algebraicity of Weil classes on split abelian eightfolds' invalidated a cancellation argument used by two dependent papers, 'Algebraicity of Kuga-Satake Correspondences for K3 Surfaces' and 'The rational Hodge conjecture for products of K3 surfaces'; revised 14 others with proof repairs and corrected statements; and added formalizations, bringing the share of top-line results formalized, by its own count, to 300 of 719. At commit fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb the manuscript map links Lean from 242 families. External reports now include MathVet's model-assisted, pre-referee statement readings and Comparator records, and Carlos Toledo's cert-machine kernel checks and family-level statement readings. For example, for family 049, MathVet reports Comparator acceptance of AbhyankarSathaye on 2026-10-08 at OpenAI commit adc7f1241b42e322a6451854ab7e4b4c146bf78a; cert-machine reports Lean and nanoda acceptance on 2026-10-10 at fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb. These are reports about that individual challenge, not kernel verdicts on the whole paper or collection.

Novelty check

Read the README, the manuscript map CONTENTS.md and the Lean catalogue lean/formalization.yaml at commit adc7f1241b42e322a6451854ab7e4b4c146bf78a on 2026-10-07. Novelty was not checked family by family, which at this scale is beyond one entry; the families split out as their own entries carry their own checks, starting with the quasi-Riemann hypothesis and Kakeya, then Artin's primitive root conjecture, modularity over imaginary quadratic fields, the Borel conjecture in dimension four, the hyperinvariant subspace problem, the O(3)O(3) model, Barnette's conjecture, eventual categoricity and the 9/49/4 matrix multiplication bound. A Hacker News commenter cross-referenced the list against ProofAtlas's ranking of 500 open problems and counted about 90 claimed full resolutions, among them Hilbert's tenth problem over the rationals, Unique Games, Landau-Siegel zeros, Baum-Connes and Hadwiger's conjecture; that count was not reproduced here. OpenAI said on 2026-09-21, when the Advisory Group on Mathematics and AI was announced, that an internal model had resolved more than 100 open problems; the registry carries no entry for that statement, which came without manuscripts.

Caveats and known objections

Graded claimed throughout. The external sources cited here report checker outcomes and attributed statement readings, not a completed human referee review of the collection. MathVet's readings remain pre-referee, and cert-machine separates its statement readings from its kernel certificates. The cited OpenAI catalogue recorded its own review status as unchecked. The README says plainly that some of the unformalized results could have issues. Where Lean exists it covers the main result of a paper rather than every claim in it, and the Comparator statements were written by OpenAI alongside the proofs, so whether each one says what the literature means by the named conjecture is a separate question for each family. A text search of the Lean library found no sorry and no axiom declarations, which is not a build and not a statement audit. The account of provenance is not checkable: the model is unreleased, a spokesperson told Scientific American that almost every result came from a single prompt to a single agent and then that some may have taken multiple attempts, and Andrew Sutherland told the magazine that one-shot claims should be treated as unverified until the model is released and the results replicated. The Advisory Group on Mathematics and AI, hosted at the Institute for Advanced Study, recommended on 2026-09-29 that labs publish the model name, prompts, a summarized chain of thought, the time taken and the estimated compute cost, and asked labs to stop testing advanced problems on proprietary models; the release gives average compute and ten reasoning summaries, not per-result prompts or costs, and Scientific American reports OpenAI saying it is not bound by the recommendations. Terence Tao has criticized the pace of releases of this kind, while Daniel Litt argued in the same article that withholding results would harm mathematics. Autonomy is graded ai-led rather than autonomous because humans chose the 4,000 problems, aggregated the output and applied the significance filter, and because the README names exceptions to the fixed procedure without describing them. This entry is a container for the release; individual families will diverge from its grades as checks land. The first withdrawals came within a day of release, on OpenAI's own initiative after a sign error, which shows both that the unformalized results can fail and that the company is correcting them. Of the 242 families cert-machine read at commit fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb, it records 127 whose Lean statements match the headline OpenAI wrote for them, 100 narrower than that headline, 11 different from it and 4 stating only a supporting result, so a kernel acceptance and a faithful statement of the advertised claim are not the same thing, and that spread is the reason this entry stays a container. The unchecked self-assessment in the catalogue is the same one that keeps 2026-09-08-navier-stokes-forced-blowup at claimed. Two families have since been machine-checked outside the company, each carrying an independent Lean formalization registered at Palomar: the L3L^3 bound for the trilinear Hilbert transform, family 086, recorded here as 2026-10-20-trilinear-hilbert-transform-lean, and the uniform interior gradient Hölder estimate for infinity-harmonic functions, recorded as 2026-10-29-infinity-harmonic-c1alpha-lean, to which the release itself attached no Lean proof.

No independent check is recorded on this entry yet.

Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.

Or on GitHub: submit a check challenge the grade send a correction or send a pull request

History of this entry

  1. CorrectedRecorded a second family machine-checked outside the company, the interior gradient Hölder estimate for infinity-harmonic functions, which the release itself left unformalized. No grade moved. source ↗
  2. CorrectedAdded cert-machine's family-level statement-reading counts, restored the Navier-Stokes precedent for the claimed grade, and pointed at the one family now independently machine-checked. No grade moved. source ↗
  3. CorrectedDated the initial no-outside-checks assessment and added MathVet and Carlos Toledo's cert-machine reports at their recorded OpenAI revisions, distinguishing checker outcomes and statement readings from completed human refereeing without changing either grade. source ↗
  4. CorrectedRecorded OpenAI's withdrawal of three algebraic-geometry manuscripts on 2026-10-07 after a sign error, which brings the release from 722 to 719 manuscripts, and its revision of 14 others. No grade moved. source ↗
  5. AddedEntered the registry graded Claimed and AI-led.

Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.

Community discussion

Graded claimed for verification and ai-led for autonomy. What these mean.

Credit

Contributed by

Cite this entry

Plain text
whataifound.org. (2026). 719 manuscripts in 372 result families from an unreleased OpenAI model. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-06-openai-math-collection
BibTeX
@misc{whataifound-openai-2026-collection,
  title        = {719 manuscripts in 372 result families from an unreleased OpenAI model},
  author       = {{whataifound.org}},
  year         = {2026},
  howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
  note         = {Result by OpenAI. Verification: Claimed. Autonomy: AI-led.},
  url          = {https://whataifound.org/finding/2026-10-06-openai-math-collection}
}

Related findings

← All mathematics findings in the registry