Humans still are numero uno when it comes to coding — but AI is sure hot on their heels.
An AI program created by OpenAI has placed second at the AtCoder World Tour Finals Heuristic coding competition held in Japan. The first place, after a nail-biting 10 hours of coding, went to human coder Psyho, who managed to closely edge out OpenAI in the contest. OpenAI had sponsored the event.
“Humanity has prevailed (for now!). I’m completely exhausted,” Psyho posted on X after it was apparent he’d win. “I figured, I had 10h of sleep in the last 3 days and I’m barely alive,” he added.
The coding problem statement had a N*N grid, with digital “robots” on them at specific positions. Coders had to write a program to be able to guide these robots to their eventual specified destinations on the grid. To make matters more complex, users could place walls on the grid, and could move several robots simultaneously with a single move. The aim was to guide all robots to their destinations in the fewest operations possible.
This is a fiendishly complicated problem, and needs high degrees of logic, creativity, and programming skills to be able to solve. Among the human coders that went about solving it in 10 hours, OpenAI had also introduced one of their reasoning models into the mix. The AI model had the same time and the same constraints as the human competitors to solve the problem.
OpenAI got off to a fast start, reaching the top of the leaderboard early. But the model had been utilizing a greedy approach, and was not looking to optimize wall movement and other more complex techniques. At that point, human coders thought that the AI had been stuck in a local minima of sorts, and humans would eventually use the more complex approaches to beat it.
But five hours into the competition, the AI began placing walls, much to the surprise of the human participants. Around the 7th hour, Psyho, the human coder who eventually won, overtook the AI on the leaderboard. But there was a twist — in the 8th hour, the model found a better approach, and jumped back to first place. But Psyho made some significant changes to his approach, and though the model tried to match them, had to settle for second place.
Now it’s remarkable that AI systems can compete on such problems at all. OpenAI says the bot worked for 10 hours autonomously. “We ran the model fully autonomously for the 10 hour window — no human intervention; same submission/data/tools/time budget as everyone else. Watching it iterate live beside elite humans was electric,” said OpenAI researcher Andre Saraiva on X. AI agents have a tendency to go off the rails and get distracted, but this agent managed to have its eye on the goal for the full 10-hour period, and produced some great results.
This was something OpenAI has already predicted. CEO Sam Altman had said earlier this year that the world’s best programmer would be an AI by the end of 2025. The company’s Chief Product Officer Kevin Weil had made a similar prediction, saying that AI will end up being better than humans in competitive coding in 2025. And with an AI bot managing to beat all humans but one in a prominent coding competition in July 2025 shows that the predictions are very much on track.