OpenAI publishes 722 maths manuscripts from an unreleased AI model
On 6 October 2026 OpenAI published new mathematical results produced by an unreleased internal frontier model. Its GitHub collection holds 722 manuscripts in 372 families. OpenAI says the model was posed about 4,000 problems and each result used roughly three hours of ChatGPT Pro thinking. Many, but not all, proofs have Lean files that a computer can check.
The short version
- 01OpenAI published 722 manuscripts in 372 families on 6 October 2026.
- 02They were produced by an internal model that OpenAI has not released.
- 03Many, but not all, come with Lean proofs; OpenAI warns some could have issues.
- 04We have not checked the maths, and no outside review was announced.
At a glance
| Published | 6 October 2026 |
|---|---|
| Company | OpenAI |
| Size | 722 manuscripts in 372 families |
| Problems posed | About 4,000 |
| Compute per result | About 3 hours of ChatGPT Pro thinking, on average |
| Where | openai/math on GitHub |
| Status | Confirmed |
What did OpenAI publish?
OpenAI released a broad range of new mathematical results produced by an internal frontier model, in its post Sharing AI progress in mathematics. The papers sit in a public GitHub repository. The catalogue has 722 manuscripts organised into 372 families, where a family groups related papers such as a main result and alternative proofs.
How were the results made?
The repository says OpenAI evaluates its models on open research problems and widened these tests after its existing maths tests were saturated. The model was posed about 4,000 problems. On average each result used the equivalent of roughly three hours of ChatGPT Pro thinking.
Can the proofs be checked?
Many, but not all, of the manuscripts come with Lean formalizations. Lean is a programming language that lets a computer check a proof. OpenAI says it will add more as it gets them.
Some of the unformalized results could have issues.
OpenAI, in the repository README
Who advised OpenAI?
OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to share the results. It also plans to fund workshops, conferences and programs about major AI results.
Can I use the model?
No. The model is unreleased. OpenAI says it is working to release it responsibly.
Why does it matter?
If the proofs hold up, an AI model is producing research-level maths at scale. That makes checking the work the bottleneck. These are OpenAI's claims, and we have not verified the maths.
What's next?
OpenAI says it will update the repository with more Lean formalizations and keep improving how it shares results.
Questions people ask
How many manuscripts did OpenAI publish?
722, organised into 372 families, according to the repository README.
Which model made them?
An unreleased internal OpenAI model. OpenAI has not named it in the release.
Are the results proven correct?
Not independently. Many have Lean proofs, but OpenAI says some unformalized results could have issues.
Can I try the model?
No. OpenAI says it is working to release it responsibly.
Sources
- PrimaryOpenAI: Sharing AI progress in mathematics (6 Oct 2026)
- PrimaryOpenAI math repository README (GitHub)
- CoverageUnite.AI: OpenAI Releases 722 Math Manuscripts From an Unreleased AI Model
Why "Confirmed"? We label a story "Confirmed" when the company or government announced it itself. Spotted an error? Email hello@openloopnews.com and we will correct it.
