It’s not every day you see an AI system score at gold medal level in a world-class contest. At the 2025 International Olympiad in Informatics (IOI), OpenAI’s system did just that—outperforming fellow AIs and even ranking sixth overall among 330 human competitors.
The AI took part in the online track under the same five‑hour time limit and 50‑submission cap as its human rivals. This time around, OpenAI opted for a mix of general‑purpose logic models rather than designing a custom solution solely for the IOI.
OpenAI researcher Noam Brown explained that the core model was the same one that recently earned a gold medal at the International Mathematical Olympiad (IMO). Rigorous evaluation showed that the model is versatile enough to handle multiple subjects, including programming, allowing it to jump into the IOI without any tweaks.
This achievement is especially rewarding after last year’s close call when the system ended up in the 49th percentile. Back then, the team had relied on a highly customised testing strategy involving a fine‑tuned model, synthetic test cases, and manual adjustments.
This year, a simpler approach was adopted. A lightweight scaffold was used to pick the best solutions from the model’s outputs, guided by another model and a clear heuristic. This streamlined method pushed the AI into the 98th percentile. Brown even hinted that soon, the core model might outperform any support system the team builds.
