OpenAI Releases AI‑Generated Mathematics Papers on GitHub

What OpenAI released and why it matters

OpenAI announced a public GitHub repository that contains dozens of new mathematical results produced by an internal frontier model, along with Lean formalizations, reasoning summaries, and compute‑usage statistics. The release is significant because it is the first large‑scale, community‑friendly disclosure of AI‑generated proofs, and it sets a precedent for how future AI breakthroughs in science should be shared.

Repository contents: results, formalizations, and transparency data

The repository (https://github.com/openai/math) is organized into three main sections:

Preprints – PDF manuscripts that describe each result, the model’s reasoning, and citations to related work.
Lean formalizations – Machine‑checked proof scripts for a subset of the results, enabling the community to verify correctness with the Lean theorem prover.
Reasoning traces – Text files that record the step‑by‑step chain‑of‑thought the model used, plus a summary of compute spent (≈ three hours of ChatGPT Pro per result on average).

"The average result used the equivalent compute of roughly three hours of ChatGPT Pro thinking." – OpenAI release notes

The repository also includes a paper‑revision protocol that specifies how authors can update manuscripts, add citations, and request formal verification.

Highlighted breakthroughs and community reaction

High‑impact problems claimed to be solved

Unique Games Conjecture – a cornerstone of hardness‑of‑approximation theory. A proof would rewrite large parts of theoretical computer science.
Barnette's Conjecture – a long‑standing graph‑theory problem; the proof is described as “approachable with basic graph theory knowledge.”
Three‑machine unit‑job scheduling – an algorithmic result with an astronomically large exponent (O((L+2)^150020) steps) that resolves a problem open since 1979.
Hilbert’s tenth problem over ℚ, Spacetime Penrose inequality, Nonexistence of Landau–Siegel zeros, among others, appear in the list of 90 solved problems out of the top 500 open problems (see comment by @zone411).

Community sentiment

Excitement – many commenters called the release “50‑100 years of math progress” and praised the open‑source approach (e.g., @binlog, @gizmodo59).
Skepticism – concerns were raised about the correctness of the proofs, the potential for loopholes, and the need for human verification (e.g., @dualvariable, @lhk931122).
Impact on researchers – some expressed anxiety about the effect on PhD work and future publications, while others saw opportunities for new workshops and collaborations (e.g., @mr_big_bowls, @novalis78).

Formal verification in Lean

OpenAI included Lean formalizations for many of the proofs, allowing the community to run the Lean checker and confirm that the logical steps are sound. This is a concrete step toward reproducible AI‑generated mathematics and aligns with recommendations from the Advisory Group on Mathematics and Artificial Intelligence (AGMAI) at the Institute for Advanced Study.

"We are publishing the results in a GitHub repository, with protocols for paper revisions and citations… As part of our GitHub repository, we are sharing formalizations of many of the proofs in Lean…" – OpenAI announcement

The Lean files are stored under lean/ and follow the standard Mathlib conventions, making it easy for existing Lean users to import and extend the work.

Compute budget and scaling implications

Each result required roughly three hours of ChatGPT Pro compute. Assuming a typical ChatGPT Pro rate of $20 USD per hour, the total compute cost for the 90‑plus results is on the order of $5,400 – a modest expense compared with the potential scientific impact.

The scaling implication is clear: as model capabilities improve, the compute per proof is expected to drop, allowing the generation of many more results with the same budget.

Open questions and next steps

Verification backlog – The community must audit the proofs, especially the high‑profile ones (Unique Games, Hilbert‑type problems). Early independent checks (e.g., Dave Goldblatt’s Lean verification of the Riemann Zeta proof) are already emerging.
Citation and authorship standards – Commenters suggested adding human reviewers’ names to the papers to indicate a level of manual validation (see @ks2048).
Broader scientific impact – While many results are pure‑math, some (e.g., the Navier–Stokes‑related computation) hint at applications in physics and engineering. Follow‑up workshops funded by OpenAI may clarify practical relevance.
Future releases – OpenAI plans to explore community‑hosted alternatives for dissemination and to improve exposition quality, as guided by AGMAI’s best‑practice recommendations.

How to use the repository today

Clone the repo: git clone https://github.com/openai/math.git.
Browse the overview PDF (overview.pdf) for a concise list of all results and their status.
Run Lean checks: install Lean 4 and Mathlib, then execute lean --make inside the lean/ directory.
Read reasoning traces to understand the model’s chain‑of‑thought and estimate compute usage.
Contribute – follow the provided revision protocol to submit errata, additional citations, or improved formalizations.

Conclusion

OpenAI’s GitHub release represents a watershed moment for AI‑augmented mathematics: it combines open‑science distribution, formal verification, and transparent compute accounting. The community now faces the dual tasks of rigorously validating the claimed breakthroughs and integrating AI‑generated insights into the broader mathematical ecosystem.