OpenAI's Experimental Model Solves 'Gold Medal-Level' Problems at the International Mathematical Olympiad
 |
| Alexander Wei_X capture |
In 2016, many of us were stunned by the arrival of AlphaGo, a groundbreaking AI that defeated Lee Sedol, the world's top Go player, with a score of 4-1. Now, we bring you news of another leap forward for artificial intelligence.
An AI has successfully challenged the highest level of human reasoning and cognition. An experimental large language model (LLM) developed by OpenAI has solved past problems from the 2025 International Mathematical Olympiad (IMO), achieving a performance equivalent to a gold medalist.
This result was shared by OpenAI researcher Alexander Wei (@alexwei_) on his X (formerly Twitter) account on July 19th. Wei emphasized the significance of the research, stating, "This achievement demonstrates that AI has overcome a long-standing barrier in its reasoning capabilities."
The model accurately solved most of the six problems on the IMO test. While it failed to arrive at the correct answer for the final problem, P6, it showed a crucial ability to recognize its own failure and suspend its judgment. Wei argued that this behavior—the model's ability to recognize that it does not know the answer—suggests that AI may be capable of metacognition.
IMO problems are not based on simple calculations but require highly advanced logical development and complex mathematical proofs, making them exceptionally difficult. This result is therefore highly significant, as it shows an AI can autonomously structure a problem and approach it logically to arrive at a solution.
What is most remarkable is that the AI did not simply generate an answer. It evaluated its own responses, recognized when they were inaccurate or uncertain, and acknowledged its own limitations in those cases. This suggests that as AI evolves to solve more complex problems, it will not just find answers but will also be able to self-correct and refine its own thought processes.