OpenAI’s New Mathematical Proofs: Has AI Crossed a Scientific Frontier?

From the Exponential Time Hypothesis to quantum chromodynamics, three remarkable claims test both the mathematical capabilities of AI and the scientific community’s ability to evaluate them.

Editorial image generated by the author.

In theoretical computer science, some problems are believed to demand exponentially growing computation even from the best algorithms. In quantum physics, theories that explain experiments with extraordinary accuracy still lack fully rigorous mathematical foundations. And in harmonic analysis, seemingly simple questions about how waves behave have occupied leading mathematicians for decades.

On October 26, 2026, OpenAI announced that its AI models had made major breakthroughs in all three areas: a claimed refutation of the Exponential Time Hypothesis, existence proofs for important four-dimensional quantum field theories, and a proof of the local smoothing conjecture with consequences for several other longstanding mathematical problems.

If the results survive independent scrutiny, they will be substantial scientific achievements. But they also pose an immediate problem of interpretation. A machine-generated mathematical argument, a formally checked theorem, and a discovery that researchers can understand and build upon are related achievements, not interchangeable ones. The three announcements provide an unusually sharp test of the distance between them.

Field	What OpenAI claims	Evidence available on October 28
Computer science	A subexponential algorithm that refutes the Exponential Time Hypothesis	Public manuscript and Lean formalization; independent examination being organized
Quantum physics	Rigorous existence constructions for QCD and related quantum field theories	Proof materials not yet publicly released; expert assessment planned
Mathematics	Local smoothing for the Euclidean wave equation in spatial dimensions two and above	Public manuscript and Lean formalization; broader mathematical scrutiny still needed

Computer Science: Is a Fundamental Limit on Computation Wrong?

The Exponential Time Hypothesis, or ETH, arose from work by Russell Impagliazzo, Ramamohan Paturi and Francis Zane in the late 1990s. It concerns how much computation is inherently required to solve certain logical problems. Its best-known formulation concerns 3-SAT: given a collection of logical conditions involving Boolean variables, can we find an assignment of true and false values that satisfies them all?

A formula with 100 binary variables has roughly 1.27 nonillion possible assignments. That is the size of the naive search space, not proof that a computer must examine every possibility. Sophisticated algorithms already do better than brute force. ETH makes a more specific claim: even the best possible algorithms cannot solve 3-SAT in subexponential time in the worst case, measured against the number of variables.

The hypothesis reaches well beyond one logical puzzle. Researchers have used ETH as an assumption in hundreds of results about computational difficulty in graph theory, computational biology, databases and optimization. A proof that ETH is false would not erase those conditional mathematical theorems, but it would force researchers to reconsider how their conclusions apply to computation in the real world.

OpenAI claims its models produced a deterministic algorithm that counts the satisfying assignments of fixed-width Boolean formulas exactly, using subexponential time and polynomial space. An exact count also answers the simpler existence question: if the count is greater than zero, at least one satisfying assignment exists. Applied to 3-SAT, the proposed algorithm would therefore refute ETH.

The public research repository includes a mathematical manuscript and a Lean formalization. The formalization’s documented scope gives a weaker running-time bound than the sharper estimates in the manuscript. Both bounds are subexponential. That is enough for the claimed challenge to ETH, provided the formalized algorithm and its input conventions correspond to the standard complexity-theoretic problem.

Subexponential time is not polynomial time, and a theoretically superior algorithm can remain useless at every practical input size. The claim neither proves P = NP nor implies that modern encryption is suddenly vulnerable. What would change is a foundational account of why certain problems are considered hard. Whether the new ideas lead to faster usable software would require further research.

Quantum Physics: Giving the Strong Force a Rigorous Foundation

Quantum chromodynamics, or QCD, describes the strong nuclear interaction: the force that governs how quarks and gluons behave. It is part of the Standard Model of particle physics, and its predictions are supported by a substantial body of experimental evidence and numerical calculation. Yet its mathematical foundations present a curious contrast with its empirical success.

Physicists can calculate with quantum field theories to extraordinary accuracy without possessing a fully rigorous construction of the relevant interacting theories in four-dimensional spacetime. The obstacle involves controlling infinities and mathematical limiting procedures while retaining the properties that make a quantum field theory physically meaningful. Successful calculations and the proof that the complete mathematical object exists are different kinds of achievement.

OpenAI reports that its models developed rigorous existence arguments for QCD, pure Yang–Mills theory and a unified extension containing the Standard Model’s gauge fields and fermions. The claimed approach builds on work by Tadeusz Balaban in constructive quantum field theory. If established, these would help close a longstanding gap between the calculations physicists use and the mathematical foundations they would like those calculations to possess.

The Yang–Mills claim needs a precise qualification. The Clay Mathematics Institute’s Millennium Prize problem asks both for a suitable rigorous quantum Yang–Mills theory to exist and for that theory to possess a positive mass gap, an energy separation between the vacuum and its lowest excited states. OpenAI claims an existence result, not a proof of the mass gap. The combined Millennium Prize problem remains unsolved.

A mathematically consistent grand-unified extension would not establish that nature follows it: an existence proof is not experimental confirmation. The immediate value of the announced work would be foundational, providing more rigorously defined objects and potentially new methods for investigating strongly interacting quantum fields.

This is also the least externally assessable announcement. OpenAI says its systems produced the Yang–Mills existence argument in September and a QCD argument later that month, and completed an internal Lean formalization of Yang–Mills existence in October. But the materials have not been made public. A group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, is to examine the work and prepare its own account. Until the mathematical construction can be inspected, outside readers cannot judge its scope or correctness in the way they can examine the published computer science and mathematics materials.

Mathematics: Why the Behavior of Waves Matters

The third claim concerns the local smoothing conjecture, proposed by Christopher Sogge in 1991. It belongs to harmonic analysis, the mathematical study of oscillations, waves and the ways complicated functions can be decomposed into simpler components.

Waves can concentrate sharply at particular places and moments. Yet concentration measured at one instant need not describe their behavior across an interval of time. Local smoothing predicts a precise improvement when the solution of a wave equation is averaged in time rather than assessed at its most extreme moment. The intuitive statement is simple; establishing the required estimates in general dimensions has been anything but.

OpenAI claims a proof for the Euclidean wave equation in every spatial dimension of at least two, using a novel auxiliary partial differential equation. The result would have consequences beyond the original question, including major estimates in Fourier restriction theory and the Kakeya problem. A Kakeya set contains a unit line segment in every direction. Such sets may have zero ordinary volume, but the conjecture asserts that they nevertheless have full Hausdorff dimension, a more subtle measure of geometric size.

This area has already seen substantial human progress. Larry Guth, Hong Wang and Ruixiang Zhang settled the two-dimensional local smoothing case in a paper published in 2020 (first circulated in 2019); Hong Wang and Joshua Zahl proved the three-dimensional Kakeya conjecture in 2025. The claimed all-dimensional extension builds on a much longer history of human ideas and partial solutions.

The most consequential element might therefore be the method. If the auxiliary equation introduced by the AI-generated argument becomes a reusable mathematical tool, the proof could matter for problems it was not originally designed to solve. That is often how a result becomes part of a discipline rather than merely another entry in its list of solved conjectures. OpenAI has released the manuscript and a Lean formalization; its formal scope statement specifies which estimates and geometric consequences are covered. Independent researchers still need to assess how the method works and where it can be used.

What Has Actually Been Verified?

The three claims differ not just in subject but in the evidence available for examining them. That is especially important because the word proof can conceal several separate stages of scientific work.

Lean is a formal proof assistant. It can check that a theorem follows from the definitions, rules, previously established results and assumptions encoded in a formal system. This is a powerful safeguard against hidden logical gaps. But the checker does not independently decide whether the formal theorem expresses the problem mathematicians intended to solve. Assumptions, dependencies and the correspondence between the formal statement and its informal interpretation remain matters for examination. The Lean reference documentation explicitly distinguishes validating a proof from understanding what its theorem statement means.

The ETH announcement illustrates the point. Its documented Lean result has a different quantitative bound from the accompanying manuscript. This need not affect the claimed refutation, but it prevents the stronger estimates from being treated as formally established merely because the weaker one was checked. A machine-checked theorem has an exact scope, and the surrounding claims must be assessed on their own terms.

The independent computer science working group is led by Impagliazzo, Paturi and Ryan Williams. In an open statement, they explicitly warn that their participation should not be treated as certification. They want the community to reconstruct the argument, test its limitations and determine its consequences. The physics group, by contrast, is working before the underlying material has been released, creating a different tension between specialist assessment and broad access.

There is a practical reason to resist both automatic acceptance and automatic dismissal. OpenAI’s repository history records three manuscripts withdrawn after a sign error earlier in October, along with revisions to other papers. These were different results, not evidence that the three new claims are wrong. They demonstrate that producing mathematically sophisticated manuscripts at scale is not the same as producing a dependable body of accepted mathematics.

The broader problem of producing mathematical proofs faster than researchers can understand them extends beyond these announcements. Here it is a practical question: what exactly has been established, and what work remains before any of these results can reshape its field?

What Do the Three Results Tell Us About AI Research?

The most striking feature of the announcement is its range. Computational complexity asks how efficiently algorithms can operate; constructive quantum field theory asks whether certain mathematical structures exist; harmonic analysis asks for precise estimates governing waves and geometry. These are not variations of a single benchmark. Success across all three would be evidence that AI systems can contribute original arguments in substantially different mathematical settings.

It would not, by itself, establish that AI has become an autonomous scientist. The conjectures were framed by human researchers, the theories draw on extensive mathematical literatures, and the public materials do not provide a complete measure of human guidance, computational expense or reproducibility. Nor can a mathematical existence proof substitute for experiments when the question is which theory describes the physical universe.

If machine-generated arguments become a regular source of frontier results, research groups will have to devote more time to reconstruction. Someone must identify which steps do the real work, distinguish essential assumptions from conveniences, relate the argument to previous methods and investigate its consequences. This is intellectual labour, not an administrative formality after the discovery.

The three release strategies make this visible. Mathematics received an open release, computer science received an open release accompanied by an organized examination, and physics received a more restricted initial process under external experts. Each approach has advantages and costs. Immediate publication allows wider scrutiny; a selected review group may help make a difficult argument intelligible but gives some researchers earlier access than others. If such discoveries become common, publication and credit practices will need to account for the intellectual work of verification and explanation alongside the production of proofs.

From a Claimed Proof to Scientific Knowledge

The three announcements are unlikely to be judged by whether they quickly produce better software, a new particle or a commercial technology. An accepted ETH refutation would change what computer scientists can plausibly rule out; a rigorous quantum-field construction would clarify the mathematical foundations of existing physics; and a reusable local-smoothing method might open paths to other analytical results.

First, however, others must reconstruct the arguments and determine what they actually prove. The published formalizations make that work possible for the computer science and mathematics claims; the physics community must still await access to the underlying material. The uneven state of the evidence is part of the story, not a footnote to it.

Whether AI has crossed a scientific frontier will be decided less by the announcement than by what follows: the theorems other mathematicians can reconstruct, the methods they can reuse and the physical questions those constructions help them investigate.

Research status checked on 28 October 2026. The verification status may change as manuscripts are examined, revised or released.