OpenAI publishes findings from unreleased model on hundreds of math problems

OpenAI has released research findings generated by an unreleased model addressing hundreds of mathematics problems, saying the results are at different stages of verification rather than being presented as fully confirmed.

The company said the findings stem from internal work with a model it has not made publicly available, and that it chose to share the results to give outside researchers visibility into the model's mathematical reasoning capabilities ahead of any public release. OpenAI did not specify a timeline for when, or whether, the underlying model itself would be released.

Mathematical problem-solving has become a closely watched benchmark for large language models, with companies including OpenAI frequently citing performance on math and reasoning tasks as evidence of progress in model capability. Independent verification of such claims typically follows as outside researchers and institutions review the published work.

OpenAI said some of the hundreds of problems addressed in the release have been verified while others remain under review, a distinction the company emphasized to avoid overstating the reliability of the full set of results.

The release adds to a steady stream of research disclosures from major AI labs this year, as companies compete to demonstrate progress on reasoning benchmarks while facing growing calls for independent scrutiny of such claims.