OpenAI has released the largest publicly available corpus of AI-generated mathematical research to date: 722 manuscripts organised into 372 families of results, produced by an internal frontier model that the company has not yet made public. The GitHub repository, live since early October 2026, pairs many of the papers with formalisations written in Lean, the proof assistant that lets a computer check each logical step of an argument — a level of rigour that natural-language text alone can never provide.
- Scale: 722 manuscripts and 372 result families, covering number theory, algebra, geometry, topology, logic, theoretical computer science and mathematical physics.
- Verification: Many results ship with Lean formalisations, but not all — OpenAI concedes that unformalised work may contain errors.
- Cost: Each result consumed roughly three hours of ChatGPT Pro compute in extended reasoning mode, drawn from a pool of about 4,000 attempted problems.
- Oversight: The release followed consultation with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, though OpenAI did not fully adopt its recommendations.
What OpenAI actually published
The archive deliberately targets contemporary pure mathematics rather than textbook exercises: analytic number theory, commutative algebra, differential topology and algebraic geometry. Each entry bundles the Lean file that certifies it, the proof written in ordinary mathematical prose, and citations to the prior theorems it depends on. Among the reported claims are a solution to the four-dimensional Kakeya conjecture, progress toward a so-called quasi-Riemann hypothesis, and work touching the Hodge conjecture and the Unique Games Conjecture.
Context matters here. The community-built Mathlib library for Lean 4 holds roughly 200,000 theorems accumulated over years of collaborative effort. OpenAI added 722 more in a single drop, all machine-checked, all aimed at problems researchers are actively working on. The company also published around ten abridged reasoning summaries and resource statistics, giving mathematicians a rare window into how the model arrived at its conclusions.
A model that already staked a claim on Navier-Stokes
The system behind the corpus is the same one that in July 2026 produced a solution to an open problem in the Navier-Stokes equations for incompressible fluids in three dimensions — one of the Clay Mathematics Institute’s Millennium Prize Problems, carrying a one-million-dollar award. That result remains under peer review, a process that can stretch over months at this level of difficulty. The fact that the same model has since generated hundreds of additional formally verified proofs suggests consistency rather than a single fortunate output.
Compute was not trivial. At roughly three hours of extended reasoning per result, the 722 manuscripts represent more than 2,100 hours of cumulative inference. OpenAI has not disclosed the infrastructure bill, but at its own published prices the tab would run into tens of thousands of dollars.
What Lean verification can and cannot prove
This is where the enthusiasm needs tempering. Lean is essentially a spell-checker for mathematical logic: it confirms that every step compiles from the stated premises, but it cannot tell you whether the conclusion is true, novel, significant, or genuinely connected to the open problem it claims to address. Formal correctness and mathematical interest are separate questions, and several repositories items carry explicit caveats or no formalisation at all.
“We want this progress to push the frontier of human knowledge and enable further progress in mathematics.”
Scepticism has already surfaced. Tristan Buckmaster of Princeton University, a specialist in partial differential equations, posted a preliminary note on arXiv questioning the real scope of the Navier-Stokes contribution, arguing the model may be constructing formally valid arguments that sidestep the deep core of the problem. The advisory group overseeing evaluation of that result issued a similar warning: Lean confirms logical consistency, not mathematical significance.
There is also a reproducibility problem. OpenAI disclosed average compute figures but withheld the exact prompts and the model itself. After controversy around the earlier Navier-Stokes claim, its advisory group recommended releasing the model, the prompts, and per-result compute time. OpenAI published the collection without fully following that advice, which limits outside researchers’ ability to retrace how any given proof came to be.
The bigger picture
The corpus is a genuine technical milestone, and formal verification is precisely what separates it from confident-sounding fiction. Yet the mathematics community is right to insist that checking a proof is not the same as understanding a theorem. What OpenAI has accelerated is one phase of mathematical labour — the construction and validation of arguments — while leaving untouched the intuition that decides which problems are worth solving and where to look next. The Stanford AI Index 2026 found that 73% of technical specialists expect AI to improve scientific research, against just 23% of the general public. This release will move that gap, in one direction or the other. Sorting valid results from merely incremental ones across 722 manuscripts is now a research burden in its own right, and it has landed squarely on human mathematicians.