For the first time, an artificial intelligence has achieved the level of a gold medal at the International Mathematical Olympiad (IMO), a prestigious competition reserved for the brightest mathematical minds. Google DeepMind’s Gemini Deep Think AI solved five of the six proposed problems, obtaining a score of 35 points out of 42, sufficient to win gold. This achievement marks a major advance in the reasoning capabilities of artificial intelligence models, with promising implications for scientific research.
A historic performance at the IMO
The International Mathematical Olympiad, organized annually since 1959, brings together the world’s best high school mathematicians. Participants, representing more than 100 countries, must solve six complex problems in two sessions of 4.5 hours, covering disciplines such as algebra, combinatorics, geometry, and number theory. Only 8% of human competitors win a gold medal, making Gemini Deep Think‘s achievement all the more remarkable.
In 2024, Google DeepMind had already impressed by winning a silver medal with its AlphaProof and AlphaGeometry 2 models, solving four of the six problems. However, these systems required human translation of problem statements into a formal language, and their resolution took up to three days. In 2025, Gemini Deep Think took a leap forward by working directly in natural language, without human intervention, and by respecting the 4.5-hour time limit imposed on human competitors. This performance was officially validated by IMO coordinators, who described the AI’s solutions as “clear, precise, and easy to follow”.
How Gemini Deep Think works
Unlike previous models, which relied on formal languages like Lean, Gemini Deep Think uses an innovative approach called “parallel thinking”. This technique allows the AI to simultaneously explore multiple paths of reasoning, integrate the results, and produce a rigorous final answer. According to Thang Luong, senior scientist at DeepMind, this method marks a paradigm shift away from traditional approaches based on reinforcement learning with verifiable final answers.
Moreover, Gemini Deep Think was trained on a high-quality set of mathematical data, including thousands of carefully selected Olympiad proofs. This training, combined with reinforcement learning strategies for multi-step reasoning, enables the AI to produce complete and elegant proofs, often simpler than those of human competitors. For example, for a number theory problem, the AI used elementary concepts where many human participants relied on advanced theorems such as Dirichlet’s.
A revealing error
Despite its success, Gemini Deep Think did not achieve a perfect score. The only unsolved problem, considered the most difficult in the competition, concerned the minimum number of rectangles needed to cover a given space. According to Junehyuk Jung, a researcher at DeepMind, the AI started with an incorrect hypothesis, assuming the answer was greater than or equal to 10, which led it to a dead end. Only five human competitors solved this problem, underscoring its complexity.
This error illustrates a current limitation of AI models: while they excel at structured reasoning, they can still be trapped by incorrect initial hypotheses. However, Gemini Deep Think‘s ability to produce clear and verifiable proofs, even for complex problems, demonstrates its potential.
Implications for research and beyond
Gemini Deep Think‘s performance goes beyond a simple competitive victory. It demonstrates that artificial intelligence models can now rival the best human minds in tasks requiring creative and abstract reasoning. For mathematicians, this advance opens the door to AI tools capable of exploring hypotheses, testing bold approaches, and accelerating the resolution of complex conjectures.
Furthermore, this technology could transform other scientific fields, such as physics, chemistry, or engineering, where complex calculations and multi-step reasoning are necessary. Google plans to deploy a version of Gemini Deep Think to mathematicians and researchers before wider availability for Google AI Ultra subscribers.
A controversy with OpenAI
Google’s announcement was preceded by a statement from OpenAI, which claimed similar performance without officially participating in the IMO. This premature announcement, evaluated by former IMO medalists rather than official organizers, sparked criticism for breaching the embargo requested by the IMO committee. In contrast, Google DeepMind waited for official validation, an approach praised by the community for its respect of the rules and human competitors.
Towards collaboration between AI and mathematicians
This breakthrough marks a turning point in the use of artificial intelligence models for mathematical reasoning. By combining the fluidity of natural language with the rigor of formal proofs, Gemini Deep Think shows that AI can become a valuable partner for researchers. However, as DeepMind emphasized, this is just the beginning. The long-term goal is to develop systems capable of solving even more complex mathematical problems, or even discovering new knowledge.
In conclusion, Gemini Deep Think‘s gold medal at the International Mathematical Olympiad illustrates the growing potential of AI to rival humans in demanding intellectual domains. This advance, supported by innovative techniques such as parallel thinking, opens exciting prospects for scientific research. However, it also reminds us of the need for close collaboration between AI and human experts to maximize its impact.