openai/math: The Repo That Turns Mathematics Into a Build Artifact

A deep dive into the Lean-backed pipeline where model output becomes papers, proofs, and formal checks, with the failures filed right alongside the wins.

10 to 12 min read • View on GitHub • More from openai

A wide editorial scene shows manuscript pages entering a verification conveyor, passing through a comparator gate, and emerging as bound preprints and proof scrolls on catalog shelves. A few pages are visibly rejected and marked as withdrawn, showing that the system keeps its failures in view instead of hiding them.
The repository behaves less like a code dump and more like an audited publishing line. Claims become manuscripts, manuscripts become formal objects, and the comparator decides what survives.
Key Takeaways

The Math Repo That Behaves Like a CI Pipeline

The first thing to understand about openai/math is that it does not behave like a normal research repository. It behaves like a controlled pipeline. The repo ships manuscripts, Lean formalizations, reasoning traces, and metadata side by side, which means every result has to survive more than one representation before it feels real.

We're releasing a broad range of new mathematical results produced by an internal frontier model.

OpenAI, Organization · Sharing AI progress in mathematics

That is the useful shift here. The project is not saying, "look, AI can write math." It is saying, "look, AI can pass through a publishing system where claims, proofs, and formal checks all have to line up." The withdrawals matter for the same reason. They show the system is willing to mark something wrong after publication, not just celebrate the headline.

What’s Actually in the Repository

The repo is organized around multiple versions of the same result. Human-readable papers live in preprints/. Lean challenge files and metadata live in lean/ComparatorChallenges/. The formal proofs sit in lean/OAI/. There are also reasoning_traces/, plus registries such as CONTENTS.md and history.md that make the whole thing navigable.

The repo is easiest to understand as a chain of custody. A claim becomes a manuscript, the manuscript becomes a formal object, and the comparator decides whether the result is genuinely stable.

openai/math/
├─ preprints/
├─ lean/
│  ├─ ComparatorChallenges/
│  └─ OAI/
├─ reasoning_traces/
├─ CONTENTS.md
└─ history.md

The Comparator Is the Real Product

The comparator is the project’s key invention. A challenge file defines the theorem interface, often with a deliberate sorry in the template. A separate solution module has to satisfy that interface without changing the target into something easier. That distinction sounds small until you realize it blocks one of the oldest cheats in automated reasoning: quietly rewriting the question.

LayerWhat it doesWhy it matters
Challenge fileStates the theorem interfaceKeeps the target fixed
Solution moduleProvides the proofMust satisfy the original interface
ComparatorChecks the matchPrevents statement tampering
ManuscriptExplains the result in proseMakes the result readable to humans

This is where Lean changes the meaning of success. A proof is no longer a persuasive explanation. It is a compilable artifact that either checks or fails. The repo’s architecture uses that fact well. It lets prose and formalization coexist without pretending they are the same thing.

A close tabletop scene shows a theorem statement on the left with a blank marked as sorry, and a dense proof sheet on the right sliding into that blank like a key into a lock. Between them sits a narrow comparator gauge, while blurred stacks of discarded drafts sit behind the main pair.
The challenge file defines the opening. The solution file has to fit that opening exactly, which is why the comparator matters more than the prose around it.

Why Lean Changes the Meaning of “Solved”

Lean matters here because it makes correctness executable. In a normal paper, the burden is on human trust and expert review. In this repo, the proof can be checked by a machine against a fixed interface. That does not eliminate judgment, but it raises the cost of bluffing.

On average, each result used three hours of ChatGPT Pro thinking compute with that model.

OpenAI, Organization · openai/math README

That disclosure is unusually revealing. It says the system is not a lightweight demo. It is doing expensive inference, then funneling that output into a formal verification layer. The result is a different epistemology from ordinary publishing. The claim is not merely written. It is checked.

The Interesting Part Is Not Just Proof, It’s Discovery

The repo is also a search engine for mathematics, not just a certificate printer. Some results are elegant confirmations. Others are constructions or counterexamples, which often matter more because they require the model to find something specific and valid, not just restate familiar structure.

ModeWhat success looks likeWhat the system has to find
ProofA conjecture is verifiedA path through existing theory
CounterexampleA conjecture is broken cleanlyA precise object that defeats the claim
ConstructionA requested object existsA valid object with the right properties

That distinction changes how you read the repo. It is not only demonstrating that a model can finish homework. It is showing that a model can participate in discovery, where the target is not obvious in advance and the result may be more interesting because it is unexpected.

The Failure Log Makes the Project Credible

If the repository were only a gallery of successes, it would be easy to dismiss. The presence of history.md changes that. Withdrawals tell you the project is willing to surface errors after release, which is exactly what a serious verification culture should do.

Everything not kept at premise at gunpoint is going to be trained on.

Tristan Buckmaster, Professor of Mathematics, NYU · OpenAI, the Partition Principle, and Mathematics

That line captures the pressure beneath the entire project. Once mathematics becomes machine-readable at scale, the discipline’s boundary conditions start to shift. More text is available for training. More claims are available for checking. More mistakes are visible when they slip through.

How It Compares to AlphaProof and Fable

The easiest comparison is not about who is "winning." It is about product shape. DeepMind’s AlphaProof is narrower and more competition-oriented. Anthropic’s Fable is closer to targeted discovery. openai/math is broader and more archival. It looks like a large public record of a reasoning system’s mathematical output.

ProjectPrimary outputTrust mechanismWhat counts as success
openai/mathManuscripts plus Lean artifactsComparator and formal checksA result that survives publication and verification
AlphaProofCompetition-grade proofsFully formal Lean proof searchA correct solution to a constrained problem set
FableTargeted research resultsTool-augmented reasoning and formalizationA high-value discovery or counterexample

That difference matters strategically. One model is optimized for a prize problem. Another is optimized for a research conjecture. This repo is optimized for scale, traceability, and reuse. It is less a trophy cabinet than a working archive.

What This Means for Mathematical Work

The biggest implication is not that AI can now write math prose. It is that mathematical work can be versioned, checked, and partially machine-authored in a pipeline that leaves receipts. That pushes the field toward a world where the proof is not just a persuasive ending. It is an artifact that can be audited, corrected, and reused.

If that sounds incremental, it is not. Once claims arrive with formal companions and failure logs, the center of gravity changes. The interesting question stops being whether a model can sound smart. It becomes whether the system can keep its chain of custody intact when the math gets hard.