An OpenAI Model Beat Every Human at One of Coding's Hardest Contests

At the AtCoder World Tour Finals in Japan, one of the most prestigious competitive programming contests in the world, an OpenAI system beat every human finalist in an exhibition match. It solved all five problems in the Algorithm Division and finished with 8,300 points — nearly double the 4,300 of the best human, a competitor called tour1st. AtCoder had even offered a “Humanity Prevails Award” of 600,000 yen (about 3,300 euros) for any human who could beat the AI and finish first. Nobody collected it.

It wasn’t a walkover, though. Competitive programmer Psyho — who famously beat OpenAI’s model at a related AtCoder final just last year — documented the contest on X. Two problems, D and E, were unusually brutal even by AtCoder standards, and the AI chewed on problem D for about three hours before cracking it. No human solved problems C or E at all. Borys Minaiev, an ICPC world champion who works on reasoning models at OpenAI, called the result “actually pretty unexpected” and said the two hard problems were tougher than anything the team had tested on. According to him, the system is a model comparable to GPT-5.6 with a small harness for scaling up thinking time, and it competed without internet access.

What’s behind this? Competitive programming is less about typing code and more about mathematical puzzle-solving under time pressure — exactly the kind of “thinking-heavy” work that was supposed to stay hard for AI. The trajectory is what stands out: in 2024, OpenAI’s system missed a bronze-level result at a top informatics olympiad; in 2025 it finished second at the AtCoder Heuristics final, losing to Psyho; now it wins outright. OpenAI stresses these systems aren’t specially trained for each contest. Worth keeping expectations grounded, though: contest problems are self-contained, clearly specified, and machine-checkable — the opposite of most real-world software work, where requirements are vague and half the job is figuring out what’s actually wanted. Winning AtCoder doesn’t mean AI can maintain your company’s messy codebase.

What this means for you: For most people, nothing changes today — but it tells you where things are heading. The reasoning muscles that win these contests are the same ones that make AI better at multi-step planning, debugging, and math-heavy work in everyday tools. If you (or your kids) are into programming contests: the human-only leaderboard still exists, but “beating the AI” has now joined chess and Go as a nostalgic idea rather than a realistic goal.

Sources

• The Decoder: OpenAI’s AI beats every human at AtCoder
• Psyho on X: contest documentation
• The Decoder: OpenAI’s second place at the 2025 Heuristics finals