OpenAI AI beats humans at AtCoder World Tour Finals 2026

OpenAI's system solved all five Algorithm Division problems at the AtCoder World Tour Finals 2026 and finished first with 8,300 points, besting the human finalists in an exhibition "Human vs. AI" match. Two of the problems, D and E, were unusually hard and stalled the AI for hours before it completed the set.

What happened at the AtCoder World Tour Finals 2026?

OpenAI's entry finished in first place with five solved problems and 8,300 points, while the runner-up tour1st scored 4,300 points. The event included a 600,000 yen "Humanity Prevails Award" for any participant who could both beat the AI and take first place, and no human competitor solved problems C or E.

The match was an exhibition alongside the annual competition that gathers top competitive programmers in Japan. Early in the contest, no human had solved more than one problem, and the AI faced an extended challenge on the two hardest problems, D and E.

How did the AI perform and what was its setup?

Two hours into the Algorithm track, problems D and E remained unsolved despite multiple attempts; the AI cracked problem D after about three hours and solved problem E shortly after OpenAI researcher Borys Minaiev left the stream. Minaiev said the result was "pretty unexpected" and that D and E were significantly harder than any AtCoder problem the team had seen before.

Minaiev described the system as a model with a small harness to scale compute at test time and said the model itself is comparable to GPT-5.6, which "ships this Thursday". The system ran without internet access. In earlier competitions the system typically solved everything in under an hour, but this contest required extended reasoning on the hardest tasks.

How does this compare with the AI's past competition runs?

The result continues a rapid climb documented across recent contests. About a year earlier, an OpenAI model placed second at the AtCoder Heuristics World Finals 2025 after running ten hours fully autonomously and being overtaken midcontest by FakePsyho. At the International Olympiad in Informatics 2025 an OpenAI system ranked sixth, ahead of all but five of the 330 human participants, and the company says the system climbed from the 49th percentile in 2024 to the 98th percentile in 2025.

At the 2025 ICPC World Finals an OpenAI system solved all twelve problems and would have taken first place; in that run GPT-5 solved eleven of twelve problems and an experimental model cracked the hardest final problem after nine submissions. OpenAI has repeatedly emphasized these systems were not trained specifically for each competition.

Why it matters

A single AI now outscored the human finalists in a top competitive programming contest under onsite conditions, solving a full Algorithm Division set that included problems observers called exceptionally difficult. That demonstrates progress on sustained, thinking-heavy algorithmic tasks rather than only fast, pattern-matched solutions. For contest organisers and competitors, the result raises questions about match formats, evaluation, and what competitive advantage general reasoning models will bring to structured problem-solving benchmarks.

What to watch

OpenAI's next likely target is the International Olympiad in Informatics 2026 in early August, a concrete milestone to test the system against another formal contest. Also watch for wider use or disclosure of models comparable to GPT-5.6, which Minaiev said the system resembles and which the article notes "ships this Thursday."