Google has announced that its “advanced version of Gemini Deep Think” has achieved a gold-medal level performance at the International Mathematical Olympiad (IMO) 2025, a prestigious competition for young mathematicians held annually since 1959.
The IMO involves pre-university mathematicians solving six exceptionally difficult problems in algebra, combinatorics, geometry, and number theory, with medals awarded to the top half of contestants, and approximately 8% receiving a gold medal.
During last week’s IMO 2025, the advanced Gemini Deep Think model perfectly solved five out of the six problems, accumulating 35 total points, a performance considered gold-medal level. Google has made the solutions publicly available.
To achieve this, the advanced Gemini was trained using novel reinforcement learning techniques designed to enhance multi-step reasoning, problem-solving, and theorem-proving capabilities. It was also provided with a curated corpus of high-quality solutions to mathematics problems and general tips for approaching IMO problems.
This achievement marks a significant advancement compared to Google DeepMind’s performance in the 2024 IMO, where AlphaGeometry and AlphaProof secured a silver medal by solving four out of six problems and scoring 28 points. A key difference this year is that Gemini “operated end-to-end in natural language, producing rigorous mathematical proofs directly from the official problem descriptions – all within the 4.5-hour competition time limit.” In contrast, the 2024 models required problems to be translated from natural language into domain-specific languages.
Deep Think is described as an “enhanced reasoning mode” that incorporates the “latest research techniques,” such as parallel thinking. This allows the model to explore and combine multiple potential solutions simultaneously, rather than following a linear thought process.
Although the Gemini Deep Think mode was announced at I/O 2025, it is not yet widely available. Google plans to make a version of this Deep Think model accessible to a select group of “trusted testers,” including mathematicians. Following this phase, it will be integrated into Google AI Ultra, though a specific launch date for subscribers of the $250 per month tier remains unconfirmed.
