OpenAI's AI wins gold at the 2025 Math Olympiad

History was rewritten last weekend in Australia, host of the 66th International Mathematical Olympiad (IMO). But this time, it wasn't a teenage prodigy who stole the spotlight, but an experimental model from OpenAI that earned a gold medal , solving five of the six challenging problems posed by the world's most demanding pre-university competition.

The announcement was made by researcher Alexander Wei through his X (formerly Twitter) account, sparking surprise and debate. “We evaluate our models under the same rules as human competitors: no tools, no internet, just pencil, paper, and reasoning,” he explained.

The IMO is considered the pinnacle of logical and creative thinking in mathematics. The fact that an AI has achieved performance equivalent to that of the best humans on the planet is not just an anecdote: it marks a turning point in the development of large language models (LLM).

These models, trained on vast amounts of text, have demonstrated skills in writing, translation, and problem-solving. But abstract mathematical reasoning has, until now, been a barely explored frontier.

Alexander Wei emphasized: “IMO problems require sustained creative thinking. This demonstrates that we can move beyond the paradigm of clear rewards and arrive at complex logical structures that resemble the human mind.”

The model's responses were evaluated by three former IMO medalists, who validated the model's reasoning by consensus. The model scored 35 out of 42 points , the score required for gold.

This breakthrough comes shortly after Google subsidiary DeepMind managed a mere silver medal in the same competition with its AlphaProof and AlphaGeometry models. Many in the scientific community believed we were still far from seeing AI performance at the highest level… until now.

However, critical voices like that of UCLA professor Ernest Ryu pointed out that solving known problems is one thing, but true mathematics invents the unknown . “I don’t think AI will replace mathematicians anytime soon,” he said, “but it will be a powerful tool for increasing their productivity.”

One of this year's most challenging problems was combinatorics , an area where there are no fixed formulas and creative thinking is key. OpenAI's model was also able to tackle this challenge, surprising even former Olympians who helped train the AI.

"You can't transform combinatorics into algebra like you can in other fields. You have to invent, and AI did that," explained one team member.

Although many expected this achievement to be part of the upcoming ChatGPT-5 , Alexander Wei clarified that the Olympic model is experimental and its release is not planned in the short term. "We want to first understand how it got so far," he noted.

The gold obtained does not pose an immediate threat, but it does represent a powerful signal that human knowledge is entering a new era of collaboration with artificial intelligence.