Who beat OpenAI's own model at a world coding championship in Tokyo?
His name is Przemyslaw Debiak. In the contest circuit everybody just says Psyho. In July 2025, at the AtCoder World Tour Finals in Tokyo, he won the heuristic track after a ten-hour marathon. Twelve humans had been invited. OpenAI entered a custom model as well. Psyho finished first. The model finished second, about nine and a half percent behind. Sam Altman posted two words: good job, Psyho.
The task was not "write a sorting function." It was a brutal optimization problem, robots moving on a 30 by 30 grid, the kind of puzzle where there is no clean exact answer and you live or die on heuristics, taste, and last-hour patches. No libraries. No docs. No Copilot whispering in your ear. Visual Studio Code and whatever you still have in your head after three days and ten hours of sleep.
The model jumped ahead early with a simple greedy plan. Psyho built something messier and smarter, passed it, lost the lead around hour eight, then took it back with late optimizations. That last stretch is the part people in this sport actually care about. Raw search is cheap now. Knowing which ugly idea to try at hour nine is not.
The detail that makes the story sting a little
Psyho used to work at OpenAI. He is on the author list for OpenAI Five, the Dota 2 agent that beat world champions in 2019. He helped stand up the training loop, the 1v1 setup, the ugly "surgery" tricks you need when you change a model without throwing away months of work. Then he left. Years later he sat down in Tokyo and beat a custom OpenAI contest agent at the thing humans were supposed to lose first.
He is also a Mensa member, a four-time TopCoder Open Marathon champion, and a Polish puzzle champion who has represented the country at world sudoku and puzzle events. He has never really done the normal full-time career. The profile reads like someone who treated hard problems as a sport and then wandered into the lab that later built ChatGPT.
Hold on. Does one contest win mean humans are still better at coding than the models?
No, and he did not claim that. He wrote "humanity has prevailed (for now)" and then said he was barely alive. Models already sit in the top hundred of a lot of coding and math contests. This was different because it was a premier onsite heuristic final, ten hours, no crutches, against a model built for that format. One data point. A loud one. Not a law of nature.
What it does show is narrower and more interesting. At the exact point where the problem stops being "emit more code" and starts being "change the strategy because the obvious one stalled," a particular kind of human still finds extra points. Poland keeps minting that kind of human. Psyho is just the one who did it on camera, against the company he once helped build.