Google DeepMind’s Gemini AI Achieves Historic Gold Medal Performance at International Mathematical Olympiad
September 18, 2025 — Google DeepMind, a leader in artificial intelligence research, has announced a landmark achievement with its advanced version of the Gemini AI model, which has reached a gold medal standard in solving complex mathematical problems at the International Mathematical Olympiad (IMO).
The announcement comes as an advanced iteration of DeepMind’s Gemini AI, dubbed “Gemini Deep Think,” successfully solved five out of six Olympiad problems flawlessly, earning 35 points — a score considered worthy of a gold medal in this prestigious global competition.
The International Mathematical Olympiad is regarded as one of the most challenging contests for pre-college students worldwide, testing creative problem-solving skills and deep mathematical reasoning. Achieving gold-level performance signifies that Gemini Deep Think can tackle problems at a level comparable to the world’s top human mathematicians.
This milestone builds on DeepMind’s longstanding strategy of creating AI systems that advance beyond pattern recognition to exhibit problem-solving capabilities in complex domains. Historically, the company’s innovations, such as AlphaFold’s breakthrough in predicting protein folding, showcased AI’s potential to catalyze scientific discovery. Gemini 2.5 and its progeny represent the next frontier, designed to push AI applications into agentic reasoning—where machines independently navigate multifaceted problems.
How Gemini Achieved the Breakthrough
Gemini Deep Think integrates multi-modal reasoning, code generation, and iterative evaluation mechanisms, enabling it to autonomously refine solutions. The model combines creativity derived from large language models with algorithmic rigor, effectively simulating the steps human problem solvers take by exploring, verifying, and optimizing mathematical proofs and computations.
In addition to the mathematical achievement, DeepMind has highlighted the broader framework underlying Gemini, which emphasizes AI as a co-scientist or problem-solving partner. This approach encourages collaboration between AI and human experts, accelerating progress across scientific disciplines.
Implications for Science and AI Development
DeepMind’s demonstration underscores a transformative shift in AI capabilities—going beyond static knowledge retrieval to dynamic reasoning that can address open-ended scientific and technical challenges. The company envisions Gemini and future AI agents aiding researchers by automating complex algorithm design, hypothesis testing, and discovery tasks.
Pushmeet Kohli, head of Google DeepMind’s science and strategic initiatives, commented in a recent Google AI podcast that such frameworks set the stage for innovations similar to AlphaFold and its successor projects, enabling AI-driven breakthroughs accessible to a wider scientific community.
Looking Forward
With this historic accomplishment, DeepMind aims to continue refining Gemini and expand its AI’s problem-solving reach. Responsible development practices remain a priority to ensure the technology benefits society while mitigating risks.
As Gemini prepares to enter broader usage scenarios, including education, research, and practical algorithm development, the AI community keenly watches how these evolving models will redefine human-AI partnerships in intellectual domains.