AI loses in coding duel: humans win after ten hours

Contents

At the AtCoder World Tour Finals 2025, Polish developer Przemysław Dębiak defeated an AI from OpenAI. In a ten-hour competition, the former OpenAI employee narrowly prevailed against the system – a result that highlights both the current limits and the potential of artificial intelligence (AI) in programming.

Human wins against AI after ten hours

As part of a special competition entitled "Humans vs AI", Dębiak –, known under the pseudonym "Psyho", competed against a specially adapted model from OpenAI. The AI system ended up in second place, ahead of ten other human finalists. Dębiak himself was exhausted after the competition, as he had apparently hardly slept during several competitions over three days. He wrote on Platform X: "Humanity has won (for now)".

What was special about the competition was the direct confrontation between an AI model and human participants under the same technical conditions. In a ten-hour final, all participants had to solve a difficult optimization problem that did not allow for an exact solution. Instead, the aim was to find the best possible approximations in 600 minutes. The competition thus offered a rare comparison between human problem-solving ability and machine efficiency in the field of competitive programming.

A level playing field for humans and machines

All participants worked on identical hardware provided by AtCoder, a Japanese online platform for programming competitions with a global ranking. All programming languages supported by the platform were available. Multiple resubmissions were allowed, but only at five-minute intervals.

According to Arstechnica, Dębiak ended up with around 1.81 trillion points, while the OpenAI model scored around 1.65 trillion – a difference of around 9.5 percent. This put the AI in second place in a field of twelve participants. Dębiak won 500,000 yen, the equivalent of around 2,900 euros.

Progress with AI in the field of programming

OpenAI described the result as a "milestone", according to Arstechnica. The model used is comparable to "o3", a system trained for strategic thinking and long-term planning. In the past, similar models have achieved good results, but have not yet reached the podium in such a competition.

In fact, many recent studies show significant progress in AI-supported programming. According to Stanford's AI Index Report 2025, AI models were already able to solve over 70 percent of tasks in a common benchmark in 2024 – the year before, it was still less than 5 percent. AI tools such as GitHub Copilot have also long been established in the everyday lives of many developers.

(mdo)