On 26 October 2026, three weeks after its first math release, OpenAI added a paper to its public math repository titled “The Exponential Time Hypothesis is False”. If it holds, it overturns a standard assumption that a lot of complexity theory builds on: a deterministic algorithm that counts satisfying assignments of fixed-width CNF formulas in subexponential time. OpenAI says an independent group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine the proof, and that the materials went public at that group’s request.
Most coverage will stop at the headline. The repository gives you enough to go a step further yourself, without being a complexity theorist. I cloned it on 18 November 2026 and used the ETH paper as a worked example of how to check any result in the collection.
What Changed in the Repository Since Launch
The README now counts 721 manuscripts in 373 families, up from 719 after the three withdrawals on 7 October. The two additions from 26 October are the ETH paper and a proof of local smoothing for the Euclidean wave equation in every dimension from two up. A Lean formalization of the rational Hodge conjecture for CM abelian varieties arrived the same day, and its comparator setup was fixed on 29 October, which is the latest commit.
The repo includes abridged reasoning summaries for ten results, including the irrationality exponent of π and Kaplansky’s direct-finiteness conjecture in characteristic two. The README still says about 42 percent of top-line results are formalized in Lean, and still describes the source as an unreleased internal model using about three hours of ChatGPT Pro thinking per result.

Step One: Find the Paper and Its Status
Every manuscript has its own folder under preprints/ with a README, the PDF and sometimes the TeX source. The ETH folder holds paper.pdf plus a build/ directory with the full TeX split into circuits, contraction and reduction parts. Check history.md before you read anything, because that’s where withdrawals and repairs are logged. The three papers pulled on 7 October each carry a notice explaining the gap, and fourteen others were revised.
If a paper you care about isn’t in lean/formalization.yaml, then it has no formal proof yet, and the written argument is all there is. On my clone that catalogue listed 176 source papers, and the ETH paper is one of them.
Step Two: Read the Formal Statement, Not the Title
A Lean proof only checks the statement it proves. The repository separates that statement into a “comparator challenge”, a short file holding just the definitions and the theorem with its proof left as sorry, so you can read what’s being claimed without wading through the proof. There are 418 challenge files in lean/ComparatorChallenges/. The one for the ETH paper is SubexponentialSAT.lean, 220 lines, and its JSON names the theorem and the only axioms the solution may use: propext, Quot.sound and Classical.choice.
The core of the claim is the time bound.
noncomputable def runtimeScale (n : ℕ) : ℝ :=
(n : ℝ) * (Real.log (Real.log ((n : ℝ) + 16))) ^ 2 /
Real.sqrt (Real.log ((n : ℝ) + 16))
The theorem exact_count_and_sat says that for every clause width k of at least 2, one fixed machine exactly counts the satisfying assignments of any width-k formula over n variables, decides satisfiability, runs in at most a constant times a polynomial in the input size times 2 raised to a constant times that scale, and never uses more than polynomial tape space. The scale grows like n times (log log n) squared over the square root of log n, which grows more slowly than any fixed multiple of n. ETH says 3-SAT needs time 2 to the power of some fixed fraction of n, so a bound of this shape, if proved, is what refuting it would look like.
That’s the check most readers skip. Here the formal statement does match the title’s claim in shape. The machine model is a home-made multi-tape Turing machine defined in the same file, so a specialist should still confirm it’s a fair model of computation, and that’s a judgment Lean can’t make for you.
Step Three: Look for Holes in the Proof Files
The solution for this theorem lives under lean/OAI/Algorithms/SubexponentialSAT/. I counted 3,977 Lean files there, about 710,000 lines, and searched them for sorry and for new axiom declarations. Neither turned up. That’s a quick sanity check, not a verification. The real test is building the project and running the comparator tool against the challenge file, and the repo’s instructions for that use lake exe cache get and lake env comparator ComparatorChallenges/<name>.json.
I didn’t run that build. The project pins Lean 4.34.1 and pulls in about 40 dependencies, including Mathlib, which is more than this machine was set up for today. If you have a Lean environment, that’s the step that turns “no sorry found” into “the kernel accepted it”.
What a Passing Check Would and Wouldn’t Tell You
If the comparator passes, you’d know a proof of that exact Lean statement exists using only the three standard axioms. You’d still need people who know the field to agree the statement means what the title says, that the machine model is fair, and that nothing in the definitions quietly weakens the claim. That’s exactly the job the Impagliazzo, Paturi and Williams group took on. The 7 October withdrawals are a reminder that the written papers can be wrong in ways Lean doesn’t catch when a result isn’t formalized.
The repository being on GitHub rather than arXiv has a practical side too: arXiv now limits each submitter to two papers a month, which no batch of 721 could fit. For another case of a headline score that needed a closer read, see what Memento 3’s 100 on ARC-AGI-3 really shows. And if you’d rather check a proof file than a paper, the steps in how to verify a FlashProve proof follow the same logic: confirm the statement, then let a small checker do the rest.
I’d expect the ETH claim to be settled by the working group long before most readers finish the PDF. Until then, the honest summary is that the formal statement says what the title says, the proof files have no obvious gaps, and the kernel check is still waiting for someone with a Lean setup to run it.