Despite significant advancements in artificial intelligence, human contestants outshone generative Artificial Intelligence (AI) models developed by Google and OpenAI at the 2025 International Mathematical Olympiad (IMO), held recently in Queensland, Australia.
AI systems reached gold-medal-level scores in the prestigious global competition for mathematicians under 20 for the first time.
However, neither Google’s Gemini model nor OpenAI’s experimental reasoning system achieved a perfect score, unlike five human contestants who earned the maximum 42 points.
Google revealed on Monday that an advanced version of its Gemini Chatbot solved five out of six complex mathematics problems presented at the Olympiad, scoring 35 out of a possible 42 points.
The result earned the model a gold medal equivalent.
“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, as quoted by Google.
“Their solutions were astonishing in many respects. IMO graders found them to be clear, precise, and most of them easy to follow.”
Similarly, OpenAI confirmed that its experimental reasoning model scored 35 points, matching the gold-medal threshold.
“We evaluated our models on the 2025 IMO problems under the same rules as human contestants,” OpenAI researcher Alexander Wei posted on social media. “For each problem, three former IMO medalists independently graded the model’s submitted proof.”
While around 10 percent of the 641 student competitors from 112 countries earned gold-level medals, only five contestants achieved a perfect 42, something neither AI model could replicate.
This year’s breakthrough marked notable progress for Google, whose AI system earned only a silver medal-equivalent score at the 2024 IMO in Bath, United Kingdom.
It also required two to three days of computation to solve four problems. In contrast, the current Gemini model completed its answers within the official 4.5-hour competition window.
The IMO confirmed that tech companies had privately tested their closed-source AI models using this year’s actual competition problems.
However, President Dolinar noted that organisers could not verify the level of computing resources used or whether there was any human involvement in the AI submissions.
“It is very exciting to see progress in the mathematical capabilities of AI models,” Dolinar added.
