OpenAI Says Its AI Models Produced Proofs That Refute Longstanding Exponential Time Hypothesis

OpenAI said on October 27, 2026 that its AI models have produced mathematical proofs that challenge longstanding theories, in an announcement that puts machine-assisted discovery at the centre of debates about how science itself is done. The company said the work covers three areas — harmonic analysis, computational complexity and quantum field theory — and that independent academic experts are being brought in to verify the findings and assess what they mean.

The headline claim is the boldest of the three. OpenAI said its system produced a proof that refutes the Exponential Time Hypothesis, a widely used conjecture about the computational effort required to solve certain problems. If that proof holds up under independent scrutiny, it would reopen questions that researchers have treated as settled limits for years — not by delivering faster algorithms overnight, but by changing what is known to be possible in principle.

What OpenAI Announced

According to the company, the three results were announced together, but they are not the same kind of claim. The computer science result concerns the Exponential Time Hypothesis, known as ETH. The harmonic analysis result concerns the local smoothing conjecture, which OpenAI said its models proved across all dimensions, a step it said also resolves the Kakeya conjecture and Stein’s restriction conjecture in all dimensions. The physics result concerns a mathematically rigorous existence proof for certain interacting quantum field theories in four dimensions, including quantum chromodynamics, which describes how particles such as quarks and gluons interact through the strong force.

That breadth is part of why the announcement has drawn attention. It is not a single narrow calculation, but a set of claims spanning pure mathematics, theoretical computer science and theoretical physics — fields where a proof is only as good as the checking that follows it. OpenAI itself framed the announcement that way, saying it hoped the results would mark the beginning of a shift towards greater use of AI in scientific discovery, while emphasising the need for independent researchers to examine and explain the findings.

Surprisingly, our system has recently proved that the ETH is false, a result that was subsequently verified in Lean.

The Exponential Time Hypothesis, in Plain Terms

The Exponential Time Hypothesis is a conjecture about how hard certain computational problems are. In simple terms, it has been used as a working assumption about the minimum effort needed to solve a class of problems, and researchers have built further results on top of it. OpenAI noted that the hypothesis has been used in hundreds of research papers to establish limits on how efficiently computational problems can be solved, across fields such as graph theory, computational biology, databases and computational geometry.

That is why a proof that the hypothesis is false matters even before any practical application appears. If a foundational assumption falls, the results built on it must be re-examined, and questions that were treated as closed may reopen. OpenAI said the finding reopens questions about how much faster these tasks could become, and may help guide the search for better algorithms. It also cautioned, in the same announcement, that finding such algorithms and turning theoretical advances into practical improvements will require significant further research.

The distinction matters for readers. A refutation of a hypothesis about computational limits is not the same as a new, faster program that anyone can run today. It is a change in the theoretical landscape: what mathematicians and computer scientists now believe is true about a class of problems, and therefore what kinds of algorithm might be worth looking for next.

Harmonic Analysis and Quantum Field Theory

The second result, in harmonic analysis, concerns how waves and signals can be broken down and understood. OpenAI said its models produced a proof of the local smoothing conjecture across all dimensions, and that this also resolves two related conjectures, Kakeya and Stein’s restriction, in all dimensions. Those names will mean most to specialists, but the pattern is the same as with ETH: a longstanding open question, a claimed proof, and a verification process that will determine whether the claim stands.

The third result moves into theoretical physics. OpenAI said its work provides a mathematically rigorous existence proof for certain interacting quantum field theories in four dimensions, including quantum chromodynamics. The work also includes a proof concerning the existence of pure Yang-Mills theory, part of a longstanding mathematics problem. Quantum field theory is the framework physicists use to describe fundamental particles and forces, so an existence proof in this area is about the mathematical foundations of the theory, not a new particle or a new experiment.

That framing helps explain why verification looks different for each result. A computer science proof can be checked in a formal system such as Lean, where every step is machine-checked. A physics existence proof must be read, understood and tested by human experts, who must judge whether the mathematics does what it claims to do. The company said it is adopting different approaches to release the findings, depending on recommendations from external academic experts — a recognition that one release model does not fit all three kinds of mathematics.

• Three results were announced: harmonic analysis, computational complexity and quantum field theory.
• The Exponential Time Hypothesis has been used in hundreds of research papers on computational limits.
• The computer science proof was verified in Lean, according to OpenAI.
• The physics result includes an existence proof for pure Yang-Mills theory.
• OpenAI committed funding for a conference and a USD 5 million contribution to the Simons Foundation to support related research.
• Independent researchers will be free to discuss the work and publish their conclusions.

How Verification Will Work

For the computer science result, OpenAI said the group will make the natural-language proof and its Lean formalisation publicly available for independent examination. That is the strongest form of openness available for a formal proof: other researchers can inspect the argument in ordinary mathematical language and also check the machine-verified version step by step.

For the physics result, the group has decided to prepare its own account before releasing the AI-produced materials. The company said the collaborations with external experts would follow principles of transparency, recognition and independence, with participation voluntary and unpaid, and researchers free to discuss the work and publish their conclusions. It also said it did not yet know which release approach would prove most useful, and welcomed feedback on what works and what could be improved.

OpenAI also said it had committed to funding a conference to examine the implications of the computer science finding and was supporting research through a USD 5 million contribution to the Simons Foundation. Those commitments do not prove the results, but they signal that the company expects the claims to be debated in public, by researchers who are not under its control. In a field where AI agents are increasingly given real responsibilities, that kind of independent checking is becoming part of the story itself, not an afterthought.

The announcement also lands in a fast-moving AI industry. Recent weeks have seen major AI companies raise their profiles sharply, including Anthropic’s record-breaking IPO filing, and OpenAI’s own work on how its models are trained and tested, including changes to live-internet access in frontier AI training. A claim about AI producing original mathematics therefore arrives at a moment when the capabilities — and the limits — of these systems are already under close public scrutiny.

What to Watch Next

What to watch next is what the article has already set out: whether independent experts confirm the proof that the Exponential Time Hypothesis is false, whether the harmonic analysis and quantum field theory results survive the same scrutiny, and which release approach proves most useful for researchers trying to check AI-produced mathematics. The other questions are whether the theoretical advance leads, over time, to better algorithms, and whether the conference and research funding OpenAI announced help turn a company announcement into a genuinely public scientific debate. Until independent researchers have examined and published their conclusions, the results remain OpenAI’s claims — striking claims, openly offered for checking, but claims all the same.