On October 6, 2026, OpenAI openly announced large-scale mathematics results generated by its internal frontier model in an official announcement titled «Sharing AI progress in mathematics». According to the company's announcement, the results were published in a GitHub repository at github.com/openai/math under the Apache-2.0 license. Both the announcement and the repository became public on the same day — October 6, 2026. «Sharing AI progress in mathematics» is the title of the announcement on the OpenAI blog; the release was issued under that name.
Publication Contents and Scale
The repository contains 722 mathematics manuscripts grouped into 372 result families. The 722 manuscripts are combined into 372 families; the release is presented in exactly that structure. The repository is public, and the license is Apache-2.0. All published materials are in open circulation under that license. The release was announced on October 6, 2026, and the repository became public on that date. The manuscript count is 722 and the family count is 372: these two figures are the release's key quantitative indicators. Apache-2.0 is the name of the release's license; the repository address is github.com/openai/math. The manuscripts contain mathematics results; they are the product of the internal frontier model's evaluation process. 4,000 is an approximate figure: the source says «roughly 4,000».
The Model and the Evaluation Process
All of the manuscripts were created using OpenAI's as-yet-unannounced internal frontier model. During evaluation, the model was tested on roughly 4,000 open mathematics problems. According to data in the GitHub repository, each result corresponds to an average of about 3 hours of ChatGPT Pro thinking compute. The model itself was not published — only the results it produced were released into the open. The number of problems involved in the evaluation and the average compute spent per result describe the scale of the release. Roughly 4,000 open problems is the measure of the evaluation's scope. And an average of 3 hours of ChatGPT Pro thinking compute per result sets the compute cost of an individual result. The internal frontier model is described as «unannounced»: the company did not issue a separate announcement about the model, publishing only the results it created.
Formalization in the Lean Language
235 of the 372 result families include proofs formalized in the Lean language. The remaining 137 families have no such formalization yet. The ratio of formalized to unformalized families shows the current state of the release: some proofs are formalized, some are not yet. 235 families is more than half of 372. The other families have no formalization yet: OpenAI openly recorded this state of affairs.
Cautions and Limitations
OpenAI openly acknowledged that unformalized results may contain issues and noted the need for caution when working with them. The company also stated that none of the published results have yet undergone peer review in scientific journals. It was separately noted that the model itself was not released: the company openly released the results but did not announce the internal frontier model that produced them. The note about the absence of peer review means the results have not yet passed the stage of scientific validation. All the cautions are stated in the release itself — in the company's official announcement.
235 of the 372 result families include Lean formalizations; OpenAI cautions that unformalized results may contain issues and notes that none of the results are peer reviewed yet and the model itself is not released.
— OpenAI, github.com/openai/math
Advisory Group and Open-Letter Context
This release came after OpenAI established an independent Advisory Group on Mathematics and AI, operating under the Institute for Advanced Study (IAS). The group is described as independent. September 21, 2026 is the date of the TechCrunch report; October 6, 2026 is the date of the OpenAI release. According to TechCrunch's September 21, 2026 report, the announcement of the group's formation came after 25 Fields medalists signed an open letter about AI companies' race to solve famous open problems. The letter is mentioned in the September 21, 2026 TechCrunch report; that same report also announced the formation of the Advisory Group.
Timeline of Events
On September 21, 2026, the formation of the Advisory Group was announced. Fifteen days later — on October 6, 2026 — the 722-manuscript release came out. Both events took place in the fall of 2026, two weeks apart. The 15-day gap connects the two events into a single context: first the Advisory Group's formation was announced, then the release came out.
