A human programmer has beaten OpenAI in a major coding contest, but the larger story is not simply that a person won. It is that the machine finished close behind, ahead of 10 other elite human competitors, in a format built to test endurance, strategy, and repeated improvement.
Przemysław Dębiak, a Polish programmer known as “Psyho” and a former OpenAI employee, narrowly defeated OpenAI’s custom AI model in the AtCoder World Tour Finals 2025 Heuristic contest in Tokyo. The event ran for 600 minutes, and Dębiak later wrote that he was “completely exhausted.”
A 10-hour contest with one hard problem
The competition centered on a single complex optimization problem. That detail matters because this was not a quick exercise in writing a small function or answering a standard coding prompt. In heuristic programming, competitors work toward better and better answers when a perfect solution would take too long to calculate.
AtCoder, a Japanese platform that hosts competitive programming contests and maintains global rankings, organized the event. The AtCoder World Tour Finals is one of the most selective competitions in the field, inviting only the top 12 programmers worldwide based on performance during the previous year.
The Heuristic division focuses on “NP-hard” optimization problems. In practical terms, that means contestants are rewarded for finding strong approximations, clever shortcuts, and strategic improvements rather than a single clean answer that can be proven best in every case.
That made the contest a revealing test for AI coding. A model could not simply produce one solution and stop. It had to reason, revise, and keep searching for better results across a long time window.
OpenAI entered as a competitor, not just a sponsor
OpenAI participated as a sponsor and also entered an AI model in a special exhibition match titled “Humans vs AI.” The model, listed as “OpenAIAHC,” was a custom simulated reasoning model similar to o3.
The setup was designed to keep the human and AI contestants on comparable footing. All competitors, including OpenAI, used identical hardware provided by AtCoder. Contestants could use any programming language available on AtCoder, and resubmissions carried no penalty, although there was a mandatory five-minute wait between submissions.
The final standings put Dębiak first with 1,812,272,558,909 points. OpenAI’s model placed second with 1,654,675,725,406 points, a margin of roughly 9.5 percent. That second-place result still put the AI ahead of 10 other human programmers who had qualified through year-long rankings.
For Dębiak, the victory came at a physical cost. He wrote on X, “Humanity has prevailed (for now!),” while noting that he had little sleep while competing in several competitions across three days. He added: “I’m completely exhausted. … I’m barely alive.”
Why the result matters for AI coding
OpenAI described the result as a milestone for competitive programming. A company spokesperson told Ars Technica: “Models like o3 rank among the top-100 in coding/math contests, but as far as we know, this is the first top-3 placement in a premier coding/math contest.”
The spokesperson also said: “Events like AtCoder give us a way to test how well our models can reason strategically, plan over long time horizons, and improve solutions through trial and error—just like a human would.”
That framing is important. The point of the contest was not only whether an AI model could write code. It was whether it could compete in a setting where progress depends on judgment, experimentation, and persistence. The model did not win, but it performed well enough to make the comparison unavoidable.
The result also fits into a broader pattern described in the source article. AI systems have improved sharply at coding tasks in recent years. Stanford University’s 2025 AI Index Report showed that on SWE-bench, “AI systems could solve just 4.4% of coding problems in 2023—a figure that jumped to 71.7% in 2024.”
AI coding tools are already part of everyday software work. The source article notes that coding is one of the most frequent uses of chatbots from OpenAI, Anthropic, Google, and Meta. It also points to tools such as GitHub Copilot and Cursor as standard tools for many professional developers, with a 2024 GitHub survey showing that over 90 percent of developers now use AI coding tools in their workflow.
A human win with a temporary feel
Dębiak’s victory stands as a strong reminder that expert human programmers still have an edge in some of the hardest competitive settings. He found a better solution under pressure, over a long contest, with little sleep and a field of elite opponents.
At the same time, the phrase “for now” captures the tension around the result. OpenAI’s model did not merely participate as a novelty. It finished second overall in a premier coding and math contest, behind one human and ahead of 10 others.
That makes the outcome less like a final answer and more like a marker in a fast-moving race. Human programmers remain capable of unexpected approaches and high-level judgment, especially in problems where no perfect answer is available. But AI models are becoming stronger at the same iterative process that defines these contests.
Dębiak appeared surprised by the wider attention. “Honestly, the hype feels kind of bizarre,” he said on X. “Never expected so many people would be interested in programming contests.”
The interest is not hard to understand. The AtCoder result compresses a much larger question into one scoreboard: how long will the best human programmers stay ahead when AI systems can reason, test, resubmit, and improve over hours? In Tokyo, the answer was clear. A human won. But the machine was close enough that the next contest may feel different.