In a report published Oct. 30, 2026, Sam Altman defended OpenAI’s decision to release AI-generated math findings, arguing that major discoveries should reach the public while acknowledging the strain the releases place on mathematicians. The Cheeky Pint episode containing his remarks was released Oct. 29.
Altman said OpenAI had more mathematical discoveries beyond the hundreds of proofs it had already released. He also estimated the company had spent hundreds of millions of dollars on math problems, with about $10 million or more spent on its Navier–Stokes effort. These were his estimates, as reported by The Verge.
What OpenAI’s repository says about the collection
In the OpenAI repository snapshot whose latest commit is shown as Oct. 30, 2026, the README lists 721 manuscripts in 373 result families and says about 42% of top-line results are formalized in Lean. OpenAI’s math repository
| Measure | README figure | Scope |
| Manuscripts | 721 | Catalogue snapshot |
| Result families | 373 | Groups of related results |
| Top-line results formalized in Lean | About 42% | OpenAI’s stated share |
| Problems posed | Approximately 4,000 | Evaluation workload |
| Compute per result | Average of 3 hours of ChatGPT Pro thinking compute | Applies to the vast majority; OpenAI lists exceptions |
OpenAI says the vast majority of results came from the same procedure, using an unnamed internal model. The company says it posed approximately 4,000 problems, then selected results for significance and grouped them into manuscripts and families. The problem count and the manuscript count describe different stages of that process. A result family can include a principal result, companion arguments, consequences or alternative proofs.
What Lean formalization tells you
Lean is a proof assistant, a software system for expressing formal proofs. OpenAI describes the collection as raw research data rather than polished academic papers. It says results are at different verification stages, not all have Lean formalizations, and unformalized results could have issues. The 42% figure refers to top-line results, while the catalogue’s 721 figure counts manuscripts.
Reported withdrawals and the experience of encountering a solution
On Oct. 21, 2026, Quartz reported that OpenAI had pulled three papers within days of the October release after errors were found. The same outlet quoted mathematician Terence Tao on how learning a solution in advance can make exploring a problem less rich and rewarding. His view highlights a concern about the process of research and learning alongside the work of checking results.
