Humans outperformed generative AI models developed by Google and OpenAI at the 2025 International Mathematical Olympiad (IMO), despite the machines achieving gold-level scores for the first time in the competition’s history.

 

 

 

Held in Queensland, Australia, the IMO featured six challenging problems to be solved within a 4.5-hour timeframe by participants under the age of 20. While Google’s Gemini and OpenAI’s reasoning model each scored 35 out of a possible 42 points, enough for a gold medal, neither achieved a perfect score. In contrast, five human contestants earned full marks.

 

 

 

 

Google announced on Monday, July 21, that its advanced Gemini chatbot had successfully solved five of the six problems, marking a significant improvement from the previous year, when it only achieved a silver-level score at the IMO in Bath, UK.

 

 

 

“We can confirm that Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points, a gold medal score,” said IMO president Gregor Dolinar. “Their solutions were astonishing in many respects. IMO graders found them to be clear, precise and most of them easy to follow.”

 

 

 

 

 

OpenAI, developers of ChatGPT, also reported a gold-level performance, with its experimental reasoning model scoring 35 points. “We evaluated our models on the 2025 IMO problems under the same rules as human contestants,” said OpenAI researcher Alexander Wei, noting that each submission was independently graded by three former IMO medalists.

 

 

 

Google noted a significant reduction in computational time this year, with its model solving the problems within the allotted competition window. In contrast, last year’s attempt took two to three days of computation.

 

 

 

The IMO confirmed that companies privately tested closed-source AI models on the same problems attempted by 641 students from 112 countries.

 

 

 

“It is very exciting to see progress in the mathematical capabilities of AI models,” Dolinar said. However, he cautioned that the contest organisers could not verify the extent of computing resources used by the AI or whether any human assistance had been involved.