OpenAI Says Its Model Disproved the Exponential Time Hypothesis: What Is Claimed, What Is Verified

OpenAI said on October 26, 2026 that its internal models produced three new results: a disproof of the Exponential Time Hypothesis (ETH) in complexity theory, a proof of the local smoothing conjecture that settles the Kakeya conjecture in every dimension, and a mathematically rigorous existence proof for four-dimensional Yang-Mills theory and QCD. The ETH and harmonic analysis results come with public Lean files. The physics result does not yet, by choice of the expert group handling it.

This is a claim from the lab that built the system, not a peer-reviewed finding. The honest framing is the one OpenAI itself uses in the announcement: the manuscripts are "raw research data" that outside experts now need to turn into understanding. Below we separate what OpenAI states, what has a machine-checkable artifact, and what is still open.

TL;DR: the three results and their status

The three different release strategies are themselves the news. OpenAI says outside experts advised each one, and that it does not yet know which will work best.

What is the Exponential Time Hypothesis, and why would a refutation matter?

The ETH is a conjecture about how hard Boolean formula satisfiability is, a problem closely tied to P versus NP. In OpenAI's words it "has been widely believed to be true, and has been used in hundreds of papers to derive conditional computational hardness results" across graph theory, computational biology, databases and computational geometry.

That last part is why this would matter if it holds. Many papers say "this problem cannot be solved much faster, unless ETH is false." If ETH is false, those conditional lower bounds lose their foundation, and researchers get a reason to hunt for faster algorithms on problems they had written off.

It is not a claim that P equals NP, and OpenAI does not make one. A Hacker News commenter asked exactly that, wondering whether the news should shift priors toward P=NP (HN comment); the announcement itself only says the refutation "reopens questions about how much faster these tasks could become" and that practical gains "will require significant further research."

The stated timeline: the natural-language disproof was found on October 16, 2026 and formalized in Lean on October 17. The collaboration group is led by Impagliazzo and Paturi, who with Francis Zane originally proposed ETH, and Williams, who has previously refuted stronger versions of it. They asked that all raw data be made public for independent verification, and OpenAI agreed, plus a commitment to fund a conference on the implications. The group has decided to publish the natural-language proof and Lean formalization immediately.

For background on how we have treated similar complexity claims, see our fact-check of the Claude 3SUM and APSP claim and the integer multiplication Lean proof post.

Kakeya and local smoothing: the math result released as raw output

The mathematics result concerns waves. Sogge's 1991 local smoothing conjecture says the most extreme focusing of waves must be short-lived: the amount of focusing at a typical time is much smaller than the worst case. It is a foundational statement, and it implies the Kakeya conjecture and Stein's restriction conjecture.

OpenAI lays out the prior record. Guth, Wang and Zhang proved local smoothing in two spatial dimensions in 2019. Wang and Zahl proved the Kakeya conjecture in three dimensions in 2025. OpenAI says its models proved local smoothing in every dimension, and therefore Kakeya and Stein's restriction conjecture in all dimensions, via "a novel auxiliary partial differential equation."

Timeline per OpenAI: AI output on October 18, Lean formalization on October 19, then an outreach to an expert, who recommended publishing the AI output and letting researchers organize themselves, the same strategy as the October 6 release. Both the natural-language write-up and Lean proof are in the public GitHub repository.

Readers should keep the earlier release in mind. The October 6 release of 722 manuscripts came with Lean formalizations for many, but not all, proofs, and OpenAI said the average result used roughly three hours of ChatGPT Pro thinking. Our breakdown of eight headline results tracks which of them have Lean status. A Kakeya-in-all-dimensions result, if it stands, would be of a different order than most of those.

The Yang-Mills and QCD claim: the one nobody can inspect yet

OpenAI describes "a mathematically rigorous existence proof of certain interacting quantum field theories containing gauge fields and fermions in four dimensions," including QCD and a unified completion of the Standard Model, building on Balaban's constructive approach. As a subproblem, it includes existence of pure Yang-Mills theory, which is part of the Clay Mathematics Institute Millennium Prize problem on Quantum Yang-Mills Theory.

Timeline per OpenAI: natural-language Yang-Mills proof on September 9, 2026; QCD proof on September 18; Lean formalization of Yang-Mills existence finished October 16. The company says it delayed release because of "the length and complexity of the work" until the Lean formalization completed.

The shepherding group is led by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. Gross shared the 2004 Nobel Prize for asymptotic freedom, and Jaffe co-formulated the Millennium problem with Witten. The group "decided to hold off on releasing the AI-produced artifacts and instead work on writing it up themselves," and asked that communications go to a dedicated email address.

Our standing guidance on Millennium claims applies. Before this announcement, our fact-check of the seven Millennium Prize Problems concluded nothing on the list had been confirmed as solved by AI. Until the group publishes its own write-up, this result is an unverifiable-by-the-public claim from OpenAI, backed by named experts who have agreed to examine it, not by a statement that they have confirmed it. Note that the announcement says the group will "understand and explain" the result; it does not say they have validated it.

The new release process: four principles and three experiments

The least flashy part of the announcement may age best. OpenAI says its "release process continues to evolve in response to feedback from the associated academic communities." The four principles for the external collaborations:

• Transparency: collaborations are disclosed, no nondisclosure agreements, participants can discuss and publish freely.
• No academic credit: OpenAI says it does not seek credit for discoveries produced by its internal models, and inviting a researcher to steward a group does not imply they made the discovery.
• Recognition: the group gets recognition for the understanding and write-ups it produces, and helps ensure prior work is credited properly.
• Independence: participation is voluntary and unpaid, and members do not speak for OpenAI.

OpenAI also names the weakness itself: "choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information." That echoes criticism we covered, including the Association for Human Mathematics statement and the AGMAI checklist for responsible release of AI mathematics. The company also points to its $5 million contribution to a Simons Foundation initiative, covered in our Simons Foundation $15M post.

What is verified and what is not

A Lean check is strong evidence that a formal statement follows from its axioms, but not that the formal statement captures the informal claim. Our Lean 4 proof cost collapse explainer covers why that matters, and our earlier look at OpenAI's Navier-Stokes agent swarm claim shows how a claim can outrun its verification. "Checked in Lean" is a pointer to where scrutiny should go, not the end of it.

What developers and readers are saying

The Hacker News thread was small when we read it, with eight comments. One commenter, rbranson, noted that the post title buried the most striking claim, that the ETH is false (comment). Another, mlinsey, asked whether complexity theorists now consider P=NP more likely (comment). Treat these as early reactions from commenters, not expert verdicts. The expert groups named by OpenAI have not yet published analysis.

What this means for what you build

Most readers will not touch Kakeya or Yang-Mills directly. The practical stakes are elsewhere:

1. Verification is becoming the bottleneck. If a frontier model can emit a Lean-checked proof in a day, the scarce resource is expert time to read it. Teams using AI for proofs, protocol verification or smart-contract analysis should budget for review of the formal statement, not just a green check.
2. Conditional hardness results may need re-reading. If your work cites ETH to justify why something cannot be sped up, track the expert group's write-up before changing anything.
3. Release practice is being invented live. Three strategies in one announcement is a useful case study if you publish AI-generated research or code and need to decide who gets access first and who gets credit.
4. Expect more. OpenAI says it is "working to responsibly release the model" behind the October 6 results, and our advisory group coverage shows how quickly this has escalated since September.

How to check it yourself

1. Read the OpenAI announcement and the October 6 release post, which explains the repository layout and protocols for paper revisions and citations.
2. Open the openai/math repository, read the natural-language write-ups, and note which results have Lean files.
3. Install Lean 4 and build the proofs. Then read the final theorem statement and ask whether it says what the headline says.
4. Wait for the named groups. For ETH, that means Impagliazzo, Paturi and Williams; the announcement says they invite an open examination and plan a funded conference.
5. Watch for dissent. Our expert reactions post shows that earlier releases drew mixed responses from mathematicians.

Related reading

• OpenAI's 722 math manuscripts: what the repo contains
• Eight headline results and their Lean status
• The seven Millennium Prize Problems: what AI has actually solved
• Did Claude break the 3SUM conjecture? Claimed versus verified
• AGMAI's rules for releasing AI math
• Simons Foundation, OpenAI and Anthropic launch a $15M effort

Details are accurate as of October 26, 2026, the day of OpenAI's announcement. Expert analysis of these results is just beginning, so check the linked sources for updates.