Google's Gemini Deep Think in IMO 2025:

• Last year, Google used an AI composed of AlphaProof and AlphaGeometry 2 models and got four out of six questions correct, earning silver medal status.
• In 2025, Google presented a new model, Gemini Deep Think, which is more analytical and runs multiple reasoning processes in parallel. It was tuned for the IMO and got five out of six questions correct, achieving gold medal status.
• Deep Think takes more time to generate output but follows the same rules as human participants by ingesting problems as natural language.

Rigorous Proofs and Deep Think's Performance:

• The IMO presents a unique challenge as questions require critical thinking and understanding of multiple mathematical disciplines.
• Deep Think's performance shows it can solve problems with simpler math when possible and recognized an incorrect hypothesis in one question.
• Only five students got one question right, and Google's 35 points earned a gold medal as only about 8% of human participants reach that level.

Evaluation and Comparison with OpenAI:

• Google emphasizes that Deep Think went through the same evaluation as students, while OpenAI did not adhere to the established process and awarded itself a gold medal.
• Google will continue to iterate on the model and participate again next year in pursuit of a perfect score. Eventually, it will be provided to Google AI Ultra subscribers.