OpenAI math research drops 722 manuscripts from unreleased model

A research release with no model attached

OpenAI's most interesting research news this month is not a model launch. It is a large pile of mathematics. On October 6, 2026, the company posted 722 mathematical manuscripts, grouped into 372 "result families," to a public GitHub repository. It credited all of them to an internal frontier model that it has not named and does not plan to release.25 With the papers came Lean formalizations for many of the results, other supporting proof material, and ten abridged summaries of how the model reasoned its way through selected problems.25

OpenAI says it posed about 4,000 problems to the model during the evaluation. It describes the average published result as using about as much compute as three hours of ChatGPT Pro thinking.510 According to the repository's README, OpenAI widened the evaluation because the model had saturated its existing math benchmarks, and some of the outputs build on results its models produced earlier.5

The release came in the same weeks as a fast run of product launches: GPT-6 Astra, a preview of GPT-5.6 Sol, GPT-6 Sol and Luna, and GPT-6.1 Sol.11131519 Taken together, these releases show where OpenAI's research effort is going. Frontier reasoning is now being shown through research artifacts rather than leaderboard scores, while the commercial story is moving toward cost per task.

What the manuscripts actually claim

The catalogue covers a lot of ground: number theory, combinatorics, theoretical computer science and mathematical physics.8 One independent review of the repository lists several headline families. They include a proof that Catalan's constant is irrational, a claim that the irrationality exponent of π is exactly two, claimed proofs of both the symmetric and general Mahler conjectures in every dimension, counterexamples tied to Kaplansky's conjectures, and a claim that free group factors are isomorphic.4 The same review points out that the π result is narrower than it first sounds. The paper itself says it does not establish bounded continued-fraction coefficients, and it gives no effective denominator threshold.4

The ten reasoning summaries cover a selected subset. Topics include ordinary two-point correlations of multiplicative functions, NP-hardness at the basic semidefinite threshold, quasipolynomial bounds for arithmetic progressions, the Mézard–Parisi formula for diluted spin glasses, and spontaneous magnetization in the quantum Heisenberg ferromagnet.5 One commentator, relaying messages from computer scientist Dana Moshkovitz via Scott Aaronson, reports that the release includes direct proofs of optimal NP-hardness of approximation for problems such as Max-Cut that do not rely on the Unique Games Conjecture at all.7 If that holds up, it would matter well beyond pure mathematics, because approximation hardness tells engineers when looking for a better algorithm is pointless.

Two results were produced differently. A paper on a zero-free region for the Riemann zeta function was edited by humans for readability, and work on the Hodge conjecture for CM abelian varieties also departed from the standard procedure.68

Where the coverage diverges

On the core numbers, the reporting broadly agrees: 722 manuscripts, 372 families, about 4,000 problems, three hours of compute per result, and ten summaries.2810 The disagreements are about verification, and they matter.

One outlet described the release as 372 proofs, each logically checked in Lean, and framed it as including progress toward the Riemann Hypothesis.3 That reading goes beyond what OpenAI said. Other coverage reports that Lean formalizations exist for many results but not all, that verification varies from paper to paper, and that papers without formalizations may contain errors.268 OpenAI itself said the results are at different stages of verification.10 One audit pointed out that a Lean artifact sitting in a repository does not mean the manuscript has been independently accepted. That team inspected proof files statically and did not run the Lean checks itself.4

The dating and counting are also loose at the edges. One report gave a total of 377 results instead of 372, and another dated the release October 7.26 Separately, claims spread on social media that the release had moved the needle on Navier–Stokes or a "quasi-Riemann" problem. One careful analysis notes that neither claim appears in OpenAI's announcement or in the repository.2 This should be kept apart from an earlier OpenAI announcement on Navier–Stokes, which other coverage treats as a separate, heavily resourced effort.910

The most reasonable reading is that this is a curated, partly formalized research dump. It is not a set of 372 certified theorems. The repository's own caveats support that view: OpenAI says citations, exposition and presentation need more work, and that it is still looking into community-hosted alternatives to a GitHub repository it controls.8

The mathematicians' objection

The pushback has focused on access rather than correctness. Nature ran a story headlined around mathematicians being "up in arms."7 TechSpot reported objections that using a proprietary model to solve problems puts results ahead of understanding, and leaves researchers unable to examine the system that produced the proofs.9 Reasoning summaries cover only ten results, a small fraction of the catalogue.9 OpenAI has not published the prompts for each result or named the model.8

OpenAI's position is that it needs to keep evaluating its internal frontier models on mathematics and other sciences so that it can build tools to advance those fields faster.9 The company also says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to present the results.10 RuntimeWire reports that the advisory group wants more than OpenAI has provided.8

I think the critics have the stronger point on process, and OpenAI has the stronger point on substance. Checkable manuscripts with formal proof targets are a far better kind of evidence than a benchmark score.4 But a field built on reproducibility cannot fully judge a method it is not allowed to run.

Benchmarks are running out of room

The math release makes more sense next to OpenAI's benchmark tables, which are getting close to their limits. On its launch page, GPT-6 Astra is listed at 97.6% on FrontierMath Tier 4 (v2), 96.0% on GPQA Diamond, 95.0% on ARC-AGI-2, and 99.9% on ARC-AGI-3.11 The ARC Prize Foundation's Greg Kamradt said Astra beat the human action-efficiency baseline on 96% of ARC-AGI-3 levels, which he called effectively human parity.11 For comparison, a third-party roundup puts GPT-5.2 at 40.3% on FrontierMath, which shows how quickly these tests are filling up.12

When scores climb that high, a benchmark stops telling models apart. That fits OpenAI's statement that it moved to open-ended problems because its internal math evaluations had saturated.5 The gains are also uneven. Astra's own table shows it behind Claude Fable 5.1 on Humanity's Last Exam with tools, 57.2% against 65.0%.11

The numbers are also less stable than they look. Astra's Terminal-Bench Science score is 64.6% on the Astra page and 68.1% on the later GPT-6.1 Sol page.1119 One analysis of benchmark reporting shows the same named test giving very different numbers depending on version, effort setting and who ran it. It cites an AutomationBench figure that appears as 18.1%, 28.8% or 60% depending on the source.17 Robotics shows a similar gap. Astra briefly led the RoboDojo leaderboard, but it struggled with tasks that need precise coordination and has since fallen to seventh place.18

Efficiency is the new headline metric

OpenAI's launch materials now lead with cost alongside accuracy. The company says GPT-6 Sol at xhigh effort beats Claude Opus 5 on AutomationBench at 9% of its cost per task. It also says GPT-6 Luna at max effort comes close to Opus 5 on DeepSWE while costing 93% less per task.15 GPT-6.1 Sol is pitched as getting within 2.1 points of Astra on OSWorld 2.0 at about one-seventh the cost.19 The GPT-5.6 Sol preview made a similar case on security work, claiming parity with a competitor on ExploitBench using about one-third of the output tokens.13 Some of the efficiency work is being done by the models themselves. A survey on arXiv cites OpenAI's report that GPT-5.6 Sol ran hundreds of experiments on its own speculative-decoding draft model, improving token-generation efficiency by more than 15%.29

Independent work on arXiv adds caution. A study of DeepSeek-R1-Distill models found that accuracy decays roughly exponentially as problems get harder, with diminishing returns from larger models, and that bigger models did not become more token-efficient.27 Another paper found that LLM-generated optimization models that get the right answer usually take longer to solve than expert-written versions.28 A third found that models editing already-optimal code made unnecessary changes 100% of the time under standard prompts.24 Researchers studying energy use found that prompting can cut the energy consumed by generated code by as much as 50%, but results varied widely, and some prompts raised consumption instead.22

Seen alongside that research, OpenAI's average of three hours of compute per result is the figure to keep an eye on. The repository gives no token ledger, no accelerator-hours, and no dollar cost. Depending on whether "result" means family, manuscript or attempted problem, the total could be anywhere from about 1,100 to 12,000 Pro-thinking hours.4 Until OpenAI publishes a full accounting, the math release shows what the model can do. It does not show what that capability costs.