IMO gold medal for OpenAI and Google DeepMind
2025-07-19 · Publisher : Google DeepMind · Domain : Artificial Intelligence
Two general-purpose language models, one from OpenAI and one from Google DeepMind, score 35 out of 42 at the 2025 International Mathematical Olympiad, gold-medal level, writing their proofs in natural language.
What changes
- Both systems solve 5 of 6 problems and score 35 out of 42, the gold threshold.
- Conditions match those of contestants: two 4.5-hour sessions, no tools or internet, proofs written in natural language.
- Google DeepMind's result (Gemini Deep Think) is certified by IMO coordinators; OpenAI's is graded by three former medalists.
- In 2024, AlphaProof had reached silver, but with formal translation of the problems and several days of compute.
- High-level mathematical reasoning is no longer out of reach for general-purpose models.
The 2025 International Mathematical Olympiad (IMO) takes place in July in Australia. On July 19, OpenAI announces that an experimental reasoning model reached gold-medal level. On July 21, Google DeepMind in turn announces a gold medal for an advanced version of Gemini Deep Think, a result certified by the organizers. ### What happened - **OpenAI**: an experimental general-purpose model, not specialized in mathematics, solves 5 of the 6 problems (35 out of 42 points) under exam conditions: two 4.5-hour sessions, no tools or internet access, proofs in natural language. Each proof is graded by three former IMO medalists, unanimously. - **Google DeepMind**: Gemini Deep Think also scores 35 out of 42. IMO President Gregor Dolinar confirms: "Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points — a gold medal score." - **Progress in one year**: in 2024, AlphaProof and AlphaGeometry reached silver-medal level, but humans had to translate the problems into formal language and computation took two to three days. ### Why it matters Both results come from general-purpose language models, using parallel reasoning and reinforcement learning techniques, not from dedicated formal proof systems. It is a strong signal about models' ability to carry out long, rigorous reasoning, beyond knowledge tests. ### Caveats OpenAI's result was not certified by the IMO. The organizers had asked AI labs to wait one week after the closing ceremony so the spotlight would stay on the contestants; OpenAI published right after the ceremony, which drew criticism, while Google DeepMind waited for certification. Neither model was public at the time of the announcement, and solving five problems out of six does not imply equivalent reliability on research work. ### Key takeaways July 2025 marks general-purpose models reaching the level of the world's best young mathematicians on a pure reasoning test.
Links
Tags : IA, Science, Raisonnement, Mathématiques, Google DeepMind, OpenAI