OpenAI Publishes 722 Mathematics Papers but Quietly Keeps Statistics and Optimization Theory Internal

On October 6, 2026, OpenAI pushed 722 mathematics papers to a public GitHub repository, claiming at once to have made substantive progress on 372 major open problems in mathematics. According to the company, the compute consumed per result was roughly equivalent to three hours of continuous reasoning by ChatGPT Pro. This was the most concentrated output by artificial intelligence in pure mathematics, with results across many fields of the mathematical alphabet—algebra, number theory, theoretical computer science, mathematical logic, and topology.

The most closely watched result was the proof of the “quasi-Riemann Hypothesis.” The quasi-Riemann Hypothesis is a weaker version of the Riemann Hypothesis, one of the seven Millennium Prize Problems in mathematics, which carries a million-dollar prize. OpenAI’s model proved that the Riemann zeta function has no zeros in the region where the real part is greater than 7/8. Researchers informally call this route the “zeta shortcut”—the model bypassed traditional proof routes and found a shorter path through the problem. The associated proof paper runs about 199 pages.

However, just one day after the public release, OpenAI withdrew three of the results after discovering a sign error. In addition, according to reports, about 50% of the results had not undergone full mathematical checking. OpenAI subsequently coordinated follow-up work with a mathematical advisory board including Harvard mathematician Melanie Wood and others. In a statement, the board characterized the release as “the beginning, not the completion, of the human process of understanding and incorporating this work into the mathematical knowledge system,” and explicitly called on OpenAI to disclose “all prompts provided to the AI agent and the agent’s full reasoning chain.”

The Missing Fields Are the Clue

The deepest discussion sparked by the public release did not come from the published results but from the “missing list.” Daniel Paleka, a researcher at ETH Zurich, systematically compared OpenAI’s outputs using the Semantic Scholar API. The conclusion was striking: statistics, machine learning theory, optimization, and numerical analysis together account for about 20% of the overall arXiv theorem corpus, yet they make up less than 1% of the 722 papers OpenAI released.

This is not a small deviation. When recalculated using a stricter baseline—“only arXiv papers that deal with already-public problems”—the figures narrowed, but the gap remained statistically significant. In other words, this cannot be explained by saying “these fields have no public hard problems”—they clearly do; OpenAI simply did not touch them.

A natural question follows: why these fields in particular? Statistics, optimization, and machine learning theory are precisely the theoretical foundations for training large language models and optimizing inference. The core methods in these fields—convergence analysis of gradient descent, tail behavior of sampling distributions, and spectral analysis of random neural networks—directly determine the training efficiency and upper performance limits of next-generation models. Paleka wrote in his analysis that he believed the plausible explanation was that “all labs tend not to publish in these mathematical fields closest to their internal uses, in order to preserve competitive advantage.”

This is not an untestable speculation. In August 2025, Sébastien Bubeck, a former principal researcher at Microsoft Research who later joined OpenAI, publicly said on social media that he had handed an open problem from a convex optimization paper to GPT-5 Pro. The model not only solved the problem but also provided a better upper-bound proof than the original paper, and he confirmed that the reasoning process was complete and correct. This shows that OpenAI’s models already had the ability to advance core mathematics in optimization as early as 2025. That capability is nowhere to be seen in the October 2026 public list.

The Absence of Cryptography: Commercial Interest or Deliberate Control?

Beyond those four fields, the absence of modern cryptography has also sparked widespread discussion. Modern public-key cryptosystems rest on hard problems in number theory, and OpenAI’s released results this time reach deeply into number theory—including bounds on prime distribution related to the quasi-Riemann Hypothesis. Matthew Green, a cybersecurity researcher and professor of cryptography at Johns Hopkins University, wrote in related discussion: “I think we may lose public-key cryptography.”

Ben Podgursky, an independent researcher, was more direct: labs succeed only on “esoteric” mathematical problems whose solution would not give the solver an asymmetric commercial advantage; this coincidence deserves attention.

There are two interpretive paths. The first is “benign self-discipline”: if cryptography were broken by AI, global financial infrastructure would face immediate risk, so labs actively choose not to publish in order to avoid a systemic crisis; this would be responsible technology management. The second is starker: these fields are not “unstudied”; rather, the research results are sealed—the boundary of model capabilities is not the same as the boundary of publicly released content.

The two interpretations are not mutually exclusive. But either way, they point together to one conclusion: OpenAI published the parts it deemed safe to publish, not the full picture of the model’s true capabilities.

The Transparency Deficit Is Accumulating

This is not the first time OpenAI has sparked a transparency controversy in mathematics. Previously, when the company claimed a breakthrough on the Navier-Stokes problem, the mathematical community raised numerous doubts about the review process and the reliability of the results, and only afterward did OpenAI form an independent mathematical advisory board. The board’s responsibilities include assessing the significance of results, recommending dissemination methods, and advising OpenAI on academic and professional standards, with the authority to proactively issue opinions and make them public.

In a public statement after this release of 722 papers, the advisory board explicitly demanded that OpenAI disclose all prompts and reasoning chains. That demand has so far received no direct response.

From an industry perspective, the stakes in this transparency dispute far exceed academic ethics. If AI models have systematically mastered core mathematical results in optimization theory and machine learning theory, and those results are used only internally and not published, then the public research community’s judgments about AI capability progress will continue to lag behind reality. This information asymmetry will keep academics, regulators, and competitors’ estimates of actual AI capabilities persistently low.

Independent Judgment

The 722 papers are a real achievement. The proof of the quasi-Riemann Hypothesis, even as a weakened version of the Riemann Hypothesis, is substantively significant in number theory. The average compute consumption of three hours per result means this kind of mathematical production is entering an industrial cadence.

But “what was published” and “what was hidden” must be scrutinized as equally important. When a company has the ability to systematically produce high-quality mathematics, yet shows a nearly complete blank in subfields highly relevant to its own commercial interests, this is not an accidental distributional skew but a choice.

The advisory board’s call to disclose reasoning chains points to the core of the issue: the credibility of AI-generated mathematical results depends not only on whether the conclusions are correct but also on whether the reasoning process can be independently checked. A publisher that acts as both referee and player, and stays silent in the fields where it has the strongest incentive for secrecy, can only allow outside doubts to keep festering. This is a question of technological openness and of industrial competitive structure, not merely a dispute over academic etiquette.