OpenAI’s GPT-5.2 Pro does better at solving sophisticated math problems than older versions of the company’s top large language model, according to a new study by Epoch AI, a non-profit research institute.

GPT-5.2 Pro solved four problems that had been too difficult for any other AI models to solve, and of the 13 problems that any other model had previously solved, it was able to solve 11, Epoch reported.

This means GPT-5.2 Pro had solved 31% of Epoch AI’s challenges, a rise from the previous score best of 19%.

Math problems have long proven difficult for AI. Scientists have speculated that this could be because AI systems can’t recognize their own limitations, while others have surmised that the issue is that AI are focused on language and not on numbers, leading to a few stumbles along the way. 

The Epoch AI experiment has demonstrated that AI is becoming more adept at some of the trickier math issues. In the test, GPT-5.2 Pro was presented with problems from various branches of math.

Joel Hass, a professor in the department of mathematics at University of California, Davis, contributed one of the problems solved by GPT-5.2 Pro. He told Epoch AI he was impressed with the way it cracked his topological challenge.  “GPT-5.2 Pro solved the problem with correct reasoning. Notably it was able to recognize the specific geometry of a surface defined by a polynomial in the problem statement,” he said.

Number theorist Ken Ono of the University of Virginia contributed another of the problems. He said that the AI model had “understood the essential theoretical trick and executed the necessary computations” to solve it, but added, “If it was a PhD student I would award only 6/10 for rigor due to missing details.”