Archive · 7 Oct 2026
OpenAI Model Solves 372 Math Problems, Altman Hails Discovery
OpenAI says one of its unreleased AI models solved 372 difficult math problems.
The problems covered several areas of mathematics.
The company says it gave the model about 4,000 problems to work on.
Each reported result took about three hours of reasoning compute on average.
Many of the proofs were checked by a computer program called Lean.
Some results have not yet been reviewed by other experts.
OpenAI shared details on GitHub and is asking an advisory group for guidance.
Sam Altman said AI could help people make more discoveries.
The announcement does not mean every result is already accepted by mathematicians.
OpenAI said an unreleased model produced 372 mathematical results across fields including algebra, number theory, topology and logic.
The company said it presented the model with about 4,000 problems, and each result took an average of roughly three hours of reasoning compute using ChatGPT Pro.
Many proofs were checked with Lean, but OpenAI said some results have not yet undergone peer review.
OpenAI published details on GitHub and said it is consulting an independent mathematics and AI advisory group on how to share the work.
Sam Altman called the announcement the beginning of a new era of discovery; the report also notes earlier criticism over an OpenAI Navier-Stokes result.
- Who
- OpenAI and its unreleased AI model; Sam Altman commented on the announcement.
- What
- OpenAI reported that the model produced 372 mathematical results.
- When
- The announcement came days after Meta reported AI-assisted solutions to six math problems; no exact date is given.
- Where
- OpenAI published details in a GitHub repository.
- Why
- OpenAI said it wants the work to expand human knowledge and enable advances in mathematics.
OpenAI and supporters
Caveats and concerns
Scientific value and openness
OpenAI and supporters
OpenAI says the results could expand human knowledge and has published methodological details, with plans to support workshops and other programs.
Caveats and concerns
The advisory group described public disclosure as the beginning of integrating the work into mathematics, and some results have not yet undergone peer review.
Claims about AI-generated mathematics
OpenAI and supporters
Sam Altman described the results as a sign of a new era of discovery.
Caveats and concerns
The article notes earlier criticism of OpenAI's Navier-Stokes result: NYU professor Tristan Buckmaster said the company might have used his work to train the AI.
- Reported results
- 372 mathematical results
- Problems presented
- Approximately 4,000
- Average compute time
- About three hours per result using ChatGPT Pro with the unreleased model
- Fields mentioned
- Algebra, number theory, theoretical computer science, mathematical logic and topology
- Proof verification
- Many proofs were checked using Lean
- Peer review
- Some results have not yet undergone peer review
- Publication
- OpenAI shared additional details in a GitHub repository