OpenAI Unreleased Model Math Proofs: 722 Manuscripts Released

OpenAI has published a large body of mathematical work it attributes to an internal, unreleased frontier model. The release includes 722 manuscripts grouped into 372 families of results, and the company says these contain solutions to hundreds of open questions. 12 The papers are posted on GitHub. Alongside them are summaries of the model's reasoning, estimates of the compute each result consumed, and statistics on how many problems the system attempted. 2

The disclosure follows a September announcement in which OpenAI said the same unreleased model had solved the Navier-Stokes problem. The company described it as a nearly 90-year-old challenge and one of the most famous open questions in mathematics, and the claim stunned the mathematics community. 12 The new release looks like an attempt to move from a single headline-grabbing claim to a body of work that outsiders can inspect.

What OpenAI actually put out

The material has three main parts.

The manuscripts. These are 722 papers organized into 372 families of related results. 12 Grouping them into families suggests many of the papers are variations, extensions or companion pieces rather than 722 fully independent discoveries. Readers should keep that in mind when weighing the raw count.

The process data. OpenAI also released:

• summaries of the model's reasoning traces;
• compute estimates for each result;
• figures on how many problems the model attempted. 2

The attempt statistics matter most. A system that works on thousands of problems and succeeds on a few hundred tells a different story than one that succeeds on nearly everything it tries. Publishing that denominator is a meaningful step toward honest accounting.

The cost figure. OpenAI says the average result used compute roughly equal to three hours of ChatGPT Pro thinking. 2 That framing ties the work to a product users already pay for. It implies that research-grade output is not necessarily out of reach in terms of compute, though an average can hide a long tail of far more expensive results.

Why the Lean proofs are the most important detail

The least flashy part of the release may be the most consequential. OpenAI has formalized many of the proofs in Lean, a programming language in which a computer can check mathematical proofs step by step. 2

This speaks to the main worry about AI-generated mathematics: plausible-looking arguments that contain subtle errors. Human referees can take months or years to vet a hard proof. Hundreds of machine-written manuscripts arriving at once would overwhelm normal peer review.

A proof that compiles in Lean has been mechanically verified, so the logical chain holds, assuming the formal statement correctly captures the intended theorem. That last condition is a real caveat. Formalization verifies the proof. It does not confirm that the formal statement says what a mathematician thinks it says, or that the problem was as open or as important as claimed.

Where the uncertainties lie

The available reporting does not specify how many of the proofs were formalized. It says "many," not all. 2 Results without Lean versions will still depend on conventional human review.

The reporting also does not settle how independent experts view the Navier-Stokes claim, beyond noting that the original announcement stunned mathematicians. 12 A result of that profile would normally get intense, prolonged scrutiny. A public archive invites that scrutiny but does not replace it.

The model behind the work also remains unreleased. Outside researchers can examine its outputs and reasoning summaries, but they cannot query the system or reproduce its attempts. The work is transparent about its results, not about the model that produced them.

Reading the move

On balance, this looks less like a victory lap and more like OpenAI trying to make its mathematical claims auditable. Releasing failure counts, compute costs and machine-checkable proofs is the kind of evidence skeptics have asked for. It is also notably more than a press release about one famous problem would require.

The real test will play out over the coming months, as mathematicians work through the archive. They will look for weak results, mis-stated theorems and problems that turn out not to have been open at all. If a substantial share holds up, especially among the Lean-verified proofs, the release could mark a shift in how mathematics gets produced and checked. Until then, the honest framing is a large, unusually well-documented set of claims that the field can now examine itself.