Recently, a Polish programmer on fumes achieved what might soon be unachievable: defeating an OpenAI advanced AI model in a head-to-head coding competition. After completing the 10-hour marathon, he was “completely exhausted.”
In the AtCoder World Tour Finals 2025 Heuristic competition in Tokyo on Wednesday, the custom AI model was nearly defeated by programmer Przemysław Dębiak (also known as “Psyho”), a former OpenAI employee. It might be the first time an AI model faced up against the best human programmers in a big onsite world championship, according to AtCoder, a Japanese platform that holds competitive programming competitions and keeps track of world rankings. As a sponsor of the event, ChatGPT’s creator submitted an AI model in a unique demonstration battle called “Humans vs. AI.” The business won second place in spite of silicon’s unrelenting nature.
On X, Dębiak said, “Humanity has prevailed (for now!)” and mentioned that he didn’t get much sleep while participating in multiple tournaments over the course of three days. “I’m completely exhausted. … I’m barely alive.”
Participants in the competition have 600 minutes to tackle a single, challenging optimization issue. The competition is reminiscent of the American legend of John Henry, the steel-driving man who, in the 1870s, competed against a steam-powered drilling machine. Similar to Henry’s renowned struggle against industrial automation, Dębiak’s triumph signifies a human specialist testing their physical limitations to demonstrate that, in an era of developing artificial intelligence, human expertise is still important.
Both tales include grueling endurance competitions: Dębiak programmed for ten hours on little sleep, and Henry drove steel spikes for hours until his heart failed. Both victories are bittersweet; Dębiak’s admission that humanity won “for now” implies he understands this may be a fleeting victory over increasingly powerful machines, while Henry’s victory but death from the effort represents the unavoidable march of automation.
Even though Dębiak outlasted the renowned steel driver and won 500,000 yen, the AtCoder World Tour Finals tests the limits of both humans and AI models by posing difficult optimization problems with no ideal solution—only progressively better ones.
The effectiveness of AI is tested against human endurance in a coding marathon
The top 12 programmers in the world are invited to the AtCoder World Tour Finals, one of the most exclusive events in competitive programming, depending on their performance over the course of the preceding year. Optimization issues that are “NP-hard” are the focus of the Heuristic division. In programming, heuristics are methods for solving problems that, when perfect answers would be too time-consuming to compute, obtain good-enough answers through informed guesses and shortcuts.
A level playing field between human and AI competitors was ensured by restricting all competitors, including OpenAI, to identical hardware supplied by AtCoder. The contest guidelines allowed competitors to use any programming language that was available on AtCoder. There was a five-minute wait between submissions, but there was no penalty for resubmitting.
Psyho finished with 1,812,272,558,909 points, while OpenAI’s model (named “OpenAIAHC”) received 1,654,675,725,406 points, a difference of almost 9.5 percent, according to the final contest results. Overtaking ten other human programmers who had qualified through a year-long ranking process, OpenAI’s artificial entrant, a custom simulated reasoning model akin to o3, came in second overall.
The second-place result was described by OpenAI as a significant achievement for AI models in competitive programming. “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,” a spokeswoman for the business stated in the report. “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.”
Growing use of AI coding
Even if OpenAI’s analysis of the contest’s consequences may come out as overly positive, there is no denying that over the past few years, numerous AI models have significantly improved their ability to do coding jobs. On the SWE-bench, a test used to gauge coding proficiency, for instance, Stanford University’s 2025 AI Index Report revealed that “AI systems could solve just 4.4% of coding problems in 2023—a figure that jumped to 71.7% in 2024.”
A 2024 GitHub survey revealed that over 90% of developers now use AI coding tools in their workflow, despite a recent study suggesting that AI assistance may not save developers as much time as they believe. Coding is one of the most common uses of chatbots from OpenAI, Anthropic, Google, and Meta. Tools like GitHub Copilot and Cursor have become standard tools for many professional developers.
Dębiak’s victory, however, feels more like a noteworthy data point in a longer trajectory than a permanent victory as AI models continue to improve at jobs like coding. Though he might have to compete against a quicker machine the next time, this programmer survived to code another day, unlike Henry, who lost his life.
On X, Dębiak stated, “To be honest, the hype feels kind of bizarre.” “Never expected so many people would be interested in programming contests.”
That human potential to come up with novel solutions is still unique for the time being. Future AtCoder competitors, meanwhile, might find themselves competing more alongside AI—or not at all—as OpenAI and other businesses continue to improve their models.
