OpenAI Publishes 722 AI Math Proofs on GitHub
An unnamed internal frontier model cracked just under a tenth of some 4,000 problems, and OpenAI has pushed 722 manuscripts covering 372 results to GitHub.
In short
OpenAI released 722 AI-generated manuscripts covering 372 result families on GitHub on October 7, 2026, the yield from roughly 4,000 problems it handed to an internal frontier model.
At a glance
- 722 manuscripts on 372 result families, online since October 7, 2026 at the github.com/openai/math repository.
- Roughly 4,000 problems were posed; just under one in ten produced a relevant solution.
- Each solution averaged about as much compute as three hours of ChatGPT Pro mode.
- Almost every result came from one prompt to one agent; the earlier Navier-Stokes run used 10,000 agents.
- OpenAI does not name the model, and Lean formalizations exist for only part of the proofs.
OpenAI has posted 722 AI-generated manuscripts covering 372 result families to its github.com/openai/math repository, dated October 7, 2026. The company fed roughly 4,000 mathematical problems to an internal frontier model, and just under one in ten came back with something it judged worth publishing.
What is actually in the repository
This is a pile of incremental results rather than one headline theorem. The reported highlights are improvements to significant computer algorithms and progress on the Riemann Hypothesis. Many of the proofs ship with Lean formalizations that let a machine recheck each step, and further formalizations are promised. Machine checking therefore covers part of the collection, not all of it, and it supplements human review instead of replacing it.
The compute bill is disclosed this time
Each solution consumed roughly three hours of ChatGPT Pro mode compute on average. Almost all of them came out of a single prompt to a single agent. That is the sharpest contrast with the earlier Navier-Stokes result, which The Decoder reports needed 10,000 agents.
The dispute that preceded this
That Navier-Stokes work concerns one of the seven Millennium Prize Problems, which carries one million US dollars in prize money that OpenAI did not claim. Critics argued the model had access to prior work by human experts, and a separate accusation holds that OpenAI suppressed references to the contribution of a researcher employed by rival Anthropic.
For this release the company consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study in advance, and attached methodology notes together with compute-cost estimates.
Why mathematicians are pushing back
Timothy Gowers, a Fields medalist on that advisory group, cautions that within one to two decades the mathematical literature could be vastly larger with no human community left that genuinely understands it. Terence Tao argues that the training of young mathematicians in particular has to keep the human element central and restrict AI tools tightly. Twenty-five Fields medalists signed an open letter titled A Severe Misalignment of AI in Mathematics.
The objection is about purpose, not capability. In mathematics, solving a problem is a tool for understanding it, and mass-produced proofs can bury the ground that understanding grows in.
What is still unverified
OpenAI does not say which model produced the proofs. How much of the 372 result families is already formally verified is not stated in the reports available to us. Whether the results hold up is a matter for peer review that is only beginning. One further German-language report on the release sat behind a paywall while we researched this piece, so its figures could not be cross-checked.
FAQ
Where are the AI-generated math proofs from OpenAI?
In the public github.com/openai/math repository, online since October 7, 2026 with 722 manuscripts on 372 result families.
Which model generated the solutions?
OpenAI calls it an internal frontier model and gives no product name; the reports do not identify it any further.
Have the proofs been verified?
Only in part. Many ship with Lean formalizations for machine checking and more are promised, while peer review by mathematicians is still underway.
