Google’s Gemini Deep Think solves 5 of 6 Math Olympiad problems for gold

Google’s Gemini Deep Think AI model achieved gold medal status at the 2025 International Math Olympiad, correctly solving five of six competition problems while adhering to official IMO rules and time constraints. This marks a significant advancement over Google’s 2024 silver medal performance and demonstrates how specialized reasoning models can match elite human mathematical problem-solving abilities.

What you should know: Gemini Deep Think represents a major evolution in AI mathematical reasoning, processing problems in natural language without requiring expert translation.

• The model runs multiple reasoning processes in parallel, integrating and comparing results before delivering final answers.
• Unlike previous systems that needed domain-specific language conversion, Deep Think handles problems end-to-end in natural language.
• Google worked directly with IMO coordinators to ensure official grading and certification, following the same 4.5-hour time limit as human participants.

The big picture: Google’s approach contrasts sharply with competitors who bypass official validation processes, highlighting the importance of rigorous testing standards in AI development.

• OpenAI announced IMO results but didn’t work with the organization for official evaluation, instead using a panel of former participants to grade its own performance.
• “We confirmed with the IMO organization that we actually solved five perfectly,” said Thang Luong, a DeepMind senior scientist. “I think anyone who didn’t go through that process, we don’t know, they might have lost one point and gotten silver.”

How the training works: Google used new reinforcement learning techniques focused on complete mathematical reasoning rather than just correct answers.

• Traditional large language model (LLM) math training emphasized final answers, leading to “incomplete reasoning” that couldn’t show proper work.
• Deep Think was trained with “long answer” solutions to mathematical problems, providing better grounding for step-by-step problem solving.
• “With this kind of training, you can actually get robust, long-form reasoning,” Luong explained.

Impressive problem-solving approach: Deep Think demonstrated sophisticated mathematical intuition, sometimes choosing simpler solutions over advanced techniques.

• For the third problem, many human competitors applied graduate-level Dirichlet’s Theorem, using mathematics beyond the competition’s intended scope.
• “Our model actually made a brilliant observation and used only elementary number theory to create a self-contained proof of the given problem,” said Junehyuk Jung, a DeepMind researcher and Brown University professor.
• The model’s one incorrect answer involved starting from a flawed hypothesis about rectangle coverage, with only five human students solving that particular problem.

Why this matters: The International Math Olympiad presents a unique benchmark for AI capabilities, requiring critical thinking across multiple mathematical disciplines including algebra, combinatorics, geometry, and number theory.

• Only about 8 percent of human IMO participants achieve gold medal status, making Deep Think’s performance particularly significant.
• The competition tests pre-university level mathematics but demands sophisticated reasoning that challenges even the most advanced AI models.

What’s next: Google plans to make the IMO-tuned version of Deep Think available to paying subscribers while continuing development for future competitions.

• The specialized model is currently being tested by mathematicians and will eventually reach Google AI Ultra subscribers who pay $250 monthly.
• DeepMind intends to return next year seeking a perfect six-problem score at the International Math Olympiad.