Gemini Deep Think achieves gold-medal standard at the IMO

Background

The International Mathematical Olympiad is an annual competition for pre-university students. Its problems differ from those of most school contests in that they ask for a complete mathematical proof rather than a single number. Competitors write their arguments out, and human coordinators read each solution to decide how much credit it earns. Medals are assigned by score, with gold going to the strongest performers, and the mark needed for each medal band moves from year to year depending on how hard the paper turns out to be.

That format makes the Olympiad a difficult target for software. A proof has to hold all the way through, and the argument usually has to be found rather than computed, so a system cannot simply search a space of answers. For years, AI systems that attempted such problems needed help: a problem statement was first translated into a formal language — a notation a machine can check step by step — before the system could begin work.

What happened

On 21 July 2025, Google DeepMind announced that an advanced version of Gemini, using a mode called Deep Think, had achieved gold-medal standard at the 2025 International Mathematical Olympiad ^[1]. The model solved five of the six problems perfectly, scoring 35 out of a possible 42 points ^[1].

The grading was not left to the company. IMO coordinators officially graded and certified the results, applying the same criteria they use for student solutions ^[1], and IMO President Gregor Dolinar confirmed the score of 35 out of 42 points ^[1]. Certification by the competition's own graders is what makes the announcement comparable to a contest result rather than an internal benchmark score.

Key details

Two features of the run were singled out. First, the model worked end-to-end in natural language: it read the official problem descriptions and produced its proofs directly, with no translation step in between ^[1]. Second, it stayed inside the competition's 4.5-hour time limit ^[1]. Both points mark a change from the year before, when systems built for the Olympiad had required problems to be converted into formal language before they could be attempted ^[1][2]. Removing that step means the model works on the problem as written, with the same ambiguity and the same absence of hints that a human contestant faces.

Why it matters

Proof problems are open-ended in a way that numerical answers are not: there is no short list of candidate outputs to check, and a solution can be right in outline yet wrong in a single step. A result on such problems therefore says something about sustained multi-step reasoning rather than about recall or calculation.

The way the score was produced matters as much as the number. Because the coordinators graded the solutions with the criteria used for students, and because the president of the Olympiad confirmed the total, the claim can be checked against a public standard instead of resting on a company's own evaluation ^[1]. That is unusual for announcements of this kind, and it is the part that makes the result usable as a reference point.

The announcement also described what would happen next. Google DeepMind said a version of the model would go first to trusted testers, including mathematicians, before being rolled out to Google AI Ultra subscribers ^[1]. The capability described was therefore not immediately available to all users; a testing phase involving working mathematicians comes first.