OpenAI reasoning system wins gold at IOI 2025

OpenAI said an internal reasoning system had scored at gold-medal level at the 2025 International Olympiad in Informatics, competing under the same conditions as the roughly 330 human contestants: identical problems, no internet access, and a fixed time and submission limit over two days. Greg Brockman, OpenAI’s president, announced the result, reporting a score of 533.29 points against a gold-medal cutoff of 438.30 — enough to place sixth among the human field and first among the handful of AI systems entered in the competition’s separate AI track. OpenAI said the system used no IOI-specific training or engineering, in contrast to the bespoke pipeline the company had built the previous year.

The comparison to 2024 was the more striking figure. That year, OpenAI’s purpose-built o1-ioi system — which generated large numbers of candidate solutions per problem and selected among them — scored 213 points, roughly the 49th percentile and short of a medal. A year later, a general-purpose reasoning system beat that score by more than double using none of the same competition-specific machinery, echoing the pattern OpenAI and Google DeepMind had shown weeks earlier when their models reached gold-medal standard at the International Mathematical Olympiad.

Three Chinese high-school students outscored OpenAI’s system outright, a detail widely noted alongside the AI result as a reminder that the top human competitors still led. As with the IMO announcement, the result came from an experimental system rather than a public product, and OpenAI did not commit to releasing it; the episode was read chiefly as evidence that general-purpose reasoning models, without contest-specific tuning, were closing on the top tier of human competitive programmers faster than the previous year’s trend implied.