Recently, Sheryl Hsu, a researcher at OpenAI, announced on social media that their internally developed AI reasoning system stood out in the global top programming competition - the 2025 International Olympiad in Informatics (IOI), winning a gold medal. This is the first time an AI has achieved such a high level in this field, and it defeated 98% of the human competitors with an overwhelming advantage, drawing widespread attention.

image.png

In this global competition, OpenAI's AI system surpassed 330 top human competitors, with only 5 competitors scoring higher than it. Hsu emphasized that during this competition, the AI fully simulated the environment of human competitors, ensuring fairness. The AI was completely offline during the competition, unable to access the Internet, thereby eliminating the possibility of an open-book exam.

Additionally, the AI system did not call any external knowledge base; all problem-solving logic was generated internally by the model. Like human competitors, the AI had only 5 hours of competition time and 50 code submission opportunities, ensuring its performance relied entirely on its own logical reasoning and problem-solving capabilities.

image.png

Notably, OpenAI did not specifically train or fine-tune its model for the IOI competition. All of this was based on an integrated system (ensemble) of a general reasoning model. The team's only intervention was selecting the best solution among multiple solutions for submission, and connecting the model with the IOI submission API. All problem-solving processes were completed autonomously by the AI, demonstrating its strong ability in complex algorithmic problems.

Last year, OpenAI faced significant setbacks in the IOI competition, ultimately earning a bronze medal, ranking only in the 49th percentile. After a year of effort and improvements, this year's AI system soared from the 49th percentile to a gold medal, proving its remarkable progress and potential.

In recent weeks, this internal reasoning system has shown excellent performance in multiple top intellectual competitions, including the AtCoder World Final and the International Mathematical Olympiad (IMO), achieving outstanding results. This has raised great expectations for OpenAI's next-generation model.