An OpenAI system defeated every human competitor at the AtCoder World Tour Finals 2026, solving all five problems in the Algorithm Division. This exhibition match took place in Japan as part of an OpenAI-sponsored event.
A “Humanity Prevails Award” of 600,000 yen was offered to anyone who could beat the AI and finish first. No one claimed it.
Competitive programmer Psyho, known online as FakePsyho, noted on X that the onsite Algorithm track usually features problems that require heavy thinking but are easy to implement. This contest included problems D and E, which were unusually difficult even by AtCoder standards.
Two problems stumped the AI for hours
The match did not go smoothly from start to finish. Psyho documented the progress on X. Two hours in, problems D and E remained unsolved despite multiple attempts. No human competitor had solved more than one problem at that point.
OpenAI’s system finally cracked problem D after about three hours. Psyho stated the AI had moved past the stage where it either quickly finds a correct solution or is completely helpless. By the end, all five problems were solved, and OpenAI’s system sat in first place by a wide margin.
Borys Minaiev, an ICPC world champion who works on reasoning models at OpenAI, commented during the Algorithm Division livestream that the result was actually pretty unexpected. He had expected the system to solve everything, but said problems D and E were significantly harder than any AtCoder problem the team had seen before.
In tests on earlier competitions, the system typically solved everything in under an hour, similar to how it handled problems A, B, and C this time.
On the system’s architecture, Minaiev said it consists of a model with a small harness to scale compute at test time. The model itself is comparable to GPT-5.6, which ships this Thursday. The system had no internet access.
Six months ago, Minaiev said, they would not have been able to solve most of these problems.
Shortly after he left the stream, OpenAI’s system solved the final problem E.
AI wins 5 to 3
About a year ago, an OpenAI AI placed second at the AtCoder Heuristics World Finals 2025. That model ran for ten hours fully autonomously under the same conditions as the human finalists. It lost its early lead midway through and was eventually overtaken by FakePsyho. OpenAI called it the first known top-three finish by an AI model at a leading programming and math competition.
This time, OpenAI’s model took first place with 8,300 points and five solved problems. Runner-up tour1st scored 4,300 points. No human competitor solved problems C or E.
The latest step in a rapid climb
The result fits a pattern that has been building for a while. At the International Olympiad in Informatics 2025, an OpenAI system scored at gold medal level. OpenAI researcher Sheryl Hsu said the system ranked sixth, ahead of all but five of the 330 human participants.
The jump over a single year stands out. In 2024, OpenAI’s system narrowly missed a bronze medal and placed in the 49th percentile. A year later, it climbed to the 98th percentile.
At the 2025 ICPC World Finals, OpenAI’s system solved all twelve problems and would have taken first place, while Google’s Gemini 2.5 Deep Think reached gold level.
In all of these runs, OpenAI has stressed that the systems were not trained specifically for each competition. At the ICPC, for example, the setup was an ensemble of general reasoning models. GPT-5 solved eleven of the twelve problems, and an experimental model cracked the hardest final problem after nine submissions, something no human team managed.
The company’s next target is likely the International Olympiad in Informatics 2026 in early August.
What it means
For competitive programmers, the margin between human and machine is closing fast. The AI did not just match human performance; it surpassed it by solving the hardest problems that stumped the finalists. This suggests that general-purpose reasoning models are becoming capable of handling the specific, abstract logic required in elite programming contests without needing custom training for those events.
