Google DeepMind’s advanced Gemini Deep Think cracked five of six IMO 2025 problems, scoring 35 points and earning a gold medal. The model solved each challenge end‑to‑end in natural language, delivering proofs within the 4.5‑hour competition window.

The achievement follows last year’s AlphaGeometry and AlphaProof silver run, which required manual language translation and days of compute. Gemini’s new parallel‑thinking architecture lets it explore multiple solution paths simultaneously, cutting reasoning time and boosting accuracy.

Researchers trained the system with reinforcement learning on high‑quality math solutions and added targeted problem‑solving hints. IMO graders praised the clarity and rigor of the proofs, confirming compliance with official standards.

DeepMind plans to release a refined Deep Think model to trusted testers before offering it to Google AI Ultra subscribers, aiming to blend natural‑language fluency with formal verification for future research.