OpenAI AI proofs are at the center of an unusual research announcement spanning mathematics, theoretical computer science and physics. On Monday, October 26, 2026, the company said its internal models produced three research results, then opened the underlying proof materials for two of them: a local smoothing result in harmonic analysis and a disproof of the Exponential Time Hypothesis in computer science. The third claim concerns interacting quantum field theories and remains under a staged release led by external physicists.
The announcement is significant not only for the subjects it touches but for how it separates machine-generated artifacts from the human work required to understand, check and explain them. OpenAI called the model outputs raw research data rather than finished discoveries. That distinction matters because a readable proof can still contain hidden errors, while a formal proof checked by Lean can expose inconsistencies in definitions, logic or computation. The company’s three collaborations are therefore being judged by different processes: direct public access for the mathematics result, public materials plus a specialist review group for computer science, and an expert-authored account before the physics artifacts are released.
OpenAI AI proofs transform wave-equation analysis
The mathematics result addresses Sogge’s local smoothing conjecture for the Euclidean wave equation in every spatial dimension of at least two. The conjecture, proposed in 1991, describes how strongly waves can focus. Its central idea is that the most extreme concentration can persist only briefly, making typical focusing much weaker than the worst possible case. The statement had been established in two dimensions by Guth, Wang and Zhang in 2019, while Wang and Zahl completed the three-dimensional case in 2025.
OpenAI says the new result used an auxiliary partial differential equation of a novel kind. In addition to proving local smoothing, it resolves the Kakeya maximal and dimension conjectures and Stein’s spherical restriction estimates throughout the interior of the conjectured range. Those results are interconnected parts of harmonic analysis, a field concerned with how complex functions and signals can be understood through elementary waves.
The generation timeline was unusually compressed. OpenAI obtained the model output on October 18, formalized it in Lean the following day, and published both an AI-written natural-language account and the Lean proof in a public repository. The company said it consulted an expert before deciding to release the material directly. That approach differs from its two other collaborations because researchers are being invited to organize their own explanatory work rather than work inside a company-selected group.
Even a machine-checked proof is not the end of scholarly assessment. Researchers still need to determine whether the statement matches the accepted mathematical problem, whether every assumption is appropriate and whether the formalization captures the intended argument. A Lean file can check the encoded derivation, but human specialists remain essential for interpreting definitions and assessing the result’s reach. The public repository therefore turns the mathematics claim into an unusually direct test of an AI-generated theorem.
A conditional complexity cornerstone faces a counterexample
The computer science claim targets the Exponential Time Hypothesis, or ETH, a widely used conjecture about the difficulty of deciding whether a Boolean formula can be satisfied. ETH is closely tied to P versus NP and has supported conditional hardness results across graph theory, computational biology, databases and computational geometry. Because so much theory has been organized around its presumed truth, a valid refutation could force researchers to revisit the boundary between feasible and infeasible computation.
OpenAI says its model produced a deterministic algorithm that exactly counts satisfying assignments for fixed-width CNF formulas in subexponential time and polynomial space. The specific case of 3-SAT is what refutes ETH. According to the public Lean documentation, the formalization covers fixed k of at least two for a canonically encoded k-CNF formula over n variables, with sharper bounds based on the number of clauses set out in the accompanying manuscript.
The model’s natural-language disproof was produced on October 16, 2026, and formalized in Lean on October 17. Review will be guided by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, who with Francis Zane proposed the hypothesis in 1999, and Ryan Williams of MIT, who has studied algorithms and lower bounds related to it. The group requested that the raw materials be made public so the wider community could examine them at the same time as the reviewers.
Independent confirmation is not yet complete. Williams said the group was working to verify and examine the proof and determine whether it holds up. He also cautioned that the result was unlikely to produce immediate practical applications, even if confirmed, though its proposed technique might affect future algorithmic work. OpenAI likewise said converting the theoretical advance into usable improvements would require substantial additional research. The practical value of the ETH refutation therefore remains unknown.
There is also an important distinction between proving a conjecture false and showing that all of the consequences built upon it fail. Individual papers may still be correct when their assumptions are weakened or their arguments avoid the disputed case. ETH has served as a convenient condition, not an oracle for every result. Confirming the new algorithm would create new opportunities, but the size of that effect will emerge only as researchers revise their assumptions and lower bounds.
Quantum field theory claim enters expert review
The physics result is the least accessible because OpenAI has not released the full artifacts. The company says the models produced a mathematically rigorous existence proof for certain interacting quantum field theories with gauge fields and fermions in four dimensions. The claimed scope includes quantum chromodynamics, which describes the strong interactions of quarks and gluons, and a unified completion of the Standard Model.
The construction builds on Balaban’s constructive approach to quantum field theory. As a subproblem, it includes an existence proof for pure Yang-Mills theory, which concerns gluons and their self-interactions without the matter fields present in QCD. This is a major technical claim, but it is not a solution of the full Clay Mathematics Institute Millennium Prize problem, which asks for the existence and construction of a quantum Yang-Mills theory more generally.
OpenAI says a natural-language Yang-Mills proof was generated on September 9, followed by a QCD proof on September 18. Formalization of the Yang-Mills existence result was completed on October 16. The company then delayed public release until a group led by David Gross could write an account of the work. Gross shared the 2004 Nobel Prize in Physics for work on asymptotic freedom and co-founded the Clay Mathematics Institute. The group also includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov.
That team plans to publish its own explanation before the AI-produced materials are released. This arrangement gives leading specialists room to assess the argument and prepare a coherent scientific account, but it also limits immediate public scrutiny. The OpenAI announcement does not provide the physics proof materials in full, and OpenAI has not confirmed a public release date. As a result, claims about the completeness or significance of the physics proof remain subject to review by the experts preparing that account.
A new release model brings unequal access
OpenAI says the collaborations operate under four principles: transparency, no academic credit sought for discoveries produced by its internal models, recognition for the groups producing explanatory work, and independence for participants and the company. There are no nondisclosure agreements, participants may discuss work in progress, and their conclusions are their own. The company is also contributing $5 million to the Simons Foundation for programs that help researchers examine and build on AI-generated work.
The model is not neutral in practice. OpenAI itself recognized that inviting a select group first creates an information advantage and can reflect the judgment of those who chose the group. Its AGMAI mathematics advisory panel recommended a more open standard: publish outputs on scholarly repositories, include Lean-formalized proofs, explain how problems were selected, and fund researchers who undertake the work of understanding results. The panel warned against choosing people to receive output first because of unequal access, borrowed reputation and biased selection.
Public access alone cannot guarantee a level playing field. Experienced mathematicians and computer scientists are better equipped to interrogate a proof, reproduce formalizations and identify subtle mismatches than most readers. Nevertheless, immediate publication gives the community a common artifact to study, while funding and explanatory groups can support deeper interpretation. The hardest issue may be preserving the benefits of expertise without converting early access into control over the record. Whether the OpenAI AI proofs survive independent scrutiny at this pace, and whether the physics writeup arrives on the same open terms as the Lean-checked files, will set the precedent for how AI-generated research is released from here.
These OpenAI AI proofs will ultimately be judged by more than the novelty of their subjects. Verification, attribution, accessibility and the translation from model output to scientific understanding are all part of the result. The public mathematics and computer science materials create opportunities for rapid scrutiny, while the physics collaboration tests a slower model centered on domain authority. If the claims survive independent examination, they could influence several fields at once. If they fail, the same release process will reveal where model generation, formalization and human oversight broke down. Either way, OpenAI AI proofs are pushing research institutions to define what credible discovery by artificial intelligence should look like.