Google DeepMind has reached a historic milestone in the world of mathematics and artificial intelligence. Its advanced AI system, Gemini Deep Think, has officially achieved a gold-medal standard at the 2025 International Mathematical Olympiad (IMO)—one of the world’s most elite math competitions for students.
Outperforming Expectations: Gemini Deep Think’s IMO Success
Gemini Deep Think solved five out of the six complex math problems presented at the IMO, scoring an impressive 35 out of 42 points. This performance earned it a gold medal ranking—an achievement reserved for only the top 8% of human participants.
IMO President Prof. Dr. Gregor Dolinar praised the AI’s performance, noting that its solutions were “clear, precise and most of them easy to follow.” The AI’s results were evaluated and verified by official IMO graders using the same criteria applied to human contestants.
Solving Problems End-to-End in Natural Language
Unlike earlier AI models that required mathematical problems to be translated into formal programming languages, Gemini operated directly with natural language input. This allowed it to generate rigorous mathematical proofs within the 4.5-hour competition window—just like a human competitor.
Next-Level Reasoning With Deep Think Mode
Gemini’s success stems from its enhanced Deep Think mode—a specialized configuration designed for high-level cognitive tasks. Deep Think enables the system to explore multiple solution paths simultaneously instead of following a linear reasoning chain. This parallel thinking approach mirrors the flexibility of expert human mathematicians.
The model was also trained on a refined dataset of high-quality mathematical solutions and used reinforcement learning strategies that empower deeper, multi-step reasoning and problem-solving capabilities.
From Silver to Gold: A Year of Rapid Progress
Gemini’s gold-medal performance is a major leap from last year’s achievement, when DeepMind’s AlphaGeometry and AlphaProof systems earned a silver-level score by solving four out of six problems. Those models required extensive computation time and formal language translation. In contrast, Gemini achieved its results in real time and in plain language.
This progress reflects a broader evolution in DeepMind’s AI research, moving toward systems that can think more fluidly, intuitively, and autonomously across complex domains.
Building Toward a Future of AI-Powered Discovery
While this achievement is rooted in mathematics, the implications are far-reaching. By combining natural language fluency with formal reasoning, AI systems like Gemini are poised to become indispensable tools for researchers, scientists, and engineers across disciplines.
DeepMind continues to explore this intersection through its ongoing work with models like AlphaProof, AlphaGeometry, and related initiatives. These systems aim to assist in everything from theorem proving to real-world scientific discovery.
If you’re interested in how DeepMind’s AI is also enabling breakthroughs in other scientific areas, you might enjoy our deep dive into AlphaGenome, which explores how AI is revolutionizing our understanding of DNA.
Looking Ahead: What’s Next for Gemini Deep Think
DeepMind plans to roll out a version of the Deep Think model to a select group of mathematicians and researchers before launching it to Google AI Ultra subscribers. This controlled release aims to gather further feedback and refine the model’s capabilities before a wider deployment.
A Collaborative Effort and a Landmark Moment
This accomplishment wouldn’t have been possible without the contributions of a large interdisciplinary team. From model training and data curation to inference optimization and evaluation, experts across DeepMind collaborated to bring this vision to life.
With full verification from the IMO board, Gemini’s solutions are confirmed as complete and correct. While the review focuses on outcomes rather than the internal workings of the model, it marks a pivotal moment in AI’s journey toward human-level reasoning and beyond.
Gemini Deep Think’s gold-medal performance at the IMO is more than a technical triumph—it’s a glimpse into the future of AI-powered intellectual exploration.
