AI-powered chatbots from Google and OpenAI excel at the International Mathematics Competition, snatching top awards
In an unprecedented feat, artificial intelligence (AI) models developed by Google's DeepMind team and OpenAI have clinched gold medals at this year's International Mathematical Olympiad (IMO). However, it is important to note that OpenAI did not officially participate in the competition as a competitor.
Instead, OpenAI's AI model was independently tested on the official 2025 IMO problems under the same conditions as human contestants—two sessions of 4.5 hours each, no external tools, and solving the actual problem statements with full natural language proofs. The model's solutions were then graded by three former IMO gold medalists, earning a gold-medal-worthy score by solving 5 out of 6 problems, achieving 35 out of 42 points.
The gold-level performance was achieved using a new experimental reasoning model based on general-purpose AI techniques rather than task-specific training for the IMO. The key innovation involved massively scaling up "test-time compute," allowing the model to think for longer periods and run many lines of reasoning in parallel. This enhanced processing power enabled the model to solve complex, high-level mathematical reasoning problems typical of the IMO.
In contrast to previous AI approaches that often relied on specialized training data or verification methods, OpenAI's model used general-purpose reinforcement learning and test-time compute scaling to reach this breakthrough. It was not tailored specifically for IMO problems but demonstrated creative, lengthy reasoning similar to top human participants.
The IMO invites top students from across the world to participate in an exam that consists of two four-and-a-half-hour sessions over two days, with a total of six questions. This year, a total of 67 human participants out of 630 also received gold medals.
While the achievement of the AI models marks a significant step forward in the realm of AI, there is still skepticism about the results due to the poor performance of publicly available models like Gemini 2.5 Pro, Grok-4, and OpenAI o4, which scored less than the required points for a bronze medal. This raises questions as to why smarter models can't be scaled or made widely available.
It is worth noting that neither DeepMind nor OpenAI's AI models used in the competition are publicly available. The poor performance of publicly available models suggests a gap in the tools that we have access to and what a more finely-tuned model can do.
There is criticism towards OpenAI for announcing their gold-level performance over the weekend, before the official scores were posted on Monday, which goes against the IMO's request for companies not to announce their scores prematurely. Despite this, the achievement of the AI models at the IMO showcases the potential of AI in advanced mathematical research and problem-solving.
The gold-medal-worthy performance by OpenAI's AI model at the 2025 International Mathematical Olympiad (IMO) was achieved using a new experimental reasoning model based on general-purpose AI techniques. This model, which did not rely on task-specific training for the IMO, used creative, lengthy reasoning similar to top human participants. However, the announcement of the achievement over the weekend, before the official scores were posted, sparked criticism towards OpenAI. Despite this, the potential of artificial intelligence in advanced mathematical research and problem-solving has been showcased through this feat, highlighting the need for further improvements in technology and artificial-intelligence, especially in education-and-self-development and learning, to bridge the gap between accessible tools and finely-tuned models. In the future, tech giants like Google and OpenAI could focus on making such smarter AI models available to the public for broader application in various sectors, including the field of tech and beyond.
