At the AtCoder World Tour Finals 2026, an AI system from OpenAI defeated all 14 invited human elite programmers in the Algorithm Division and solved all five competition problems as the only participant. The result was achieved on July 9, 2026, as part of a specially announced “Human vs. AI” exhibition, as confirmed by the official competition page of AtCoder.
Elite Field Against an AI Without Internet Access
The AtCoder World Tour Finals are considered one of the most challenging programming tournaments in the world: only those who have qualified through regular AtCoder competitions are invited. In the Algorithm Division 2026, top programmers such as tourist, jiangly, ecnerwala, and ksun48 were among the participants. According to the competition page, an AI agent from OpenAI also participated – as an exhibition entry outside the regular ranking, but with the same seven hours of time and the same five problems. AtCoder additionally announced a “Humanity Prevails Award” of 600,000 yen for any human who beats the AI and finishes first.
According to consistent reports, OpenAI deployed a system whose capabilities are said to correspond roughly to the GPT-5.6 released one day later, supplemented by a streamlined harness for scaling computing power at query time. During the competition, the system had no access to the internet.
Five out of Five: How the AI Solved All Problems
According to consistent reports from MLQ News and OfficeChai, the OpenAI system solved the first three problems (A, B, C) within the first hour. Problem D then took about three hours, and Problem E was solved shortly thereafter. In the end, the system achieved a perfect score of 8,300 out of 8,300 possible points. The best human participant, ranked as “tour1st,” scored 4,300 points and stopped at three solved problems. No human in the entire field solved Problem C (1,500 points) or the hardest Problem E (2,500 points).
To illustrate the discrepancy between the problems, user CHOI on X noted: according to the difficulty scale commonly used in the scene by competitor E869120, a problem rated at 1,500 points already represents a challenge that researchers could work on for months – the two problems rated at 2,500 points are significantly more difficult still.
Reactions from the Competition Scene
Borys Minaiev, who according to MLQ News was actively involved with OpenAI in the competition and is a former ICPC world champion in programming, described the two later problems as “significantly harder than any AtCoder problem the team had seen before.” AtCoder founder Chokudai is quoted by OfficeChai as saying that despite ample preparation time, he was “utterly overwhelmed by AI.” Last year’s winner and competition commentator Psyho, who had narrowly kept the AI in second place in the Heuristic Division in 2025, commented on this year’s result according to OfficeChai, suggesting that the era of human superiority in this format is now over.
On July 7 and 8, 2026, an AI system from OpenAI also competed in the separately ranked Heuristic Division of the same event – an optimization rather than algorithm discipline. Psyho had pointed out in advance on X that an AI had already placed second there in 2025. Detailed, officially confirmed final results for the Heuristic Division 2026 were not available at the time of this post; according to AtCoder, interim rankings shown during the competition were explicitly considered provisional.
Assessment
Competitive programming contests like AtCoder have long been regarded in AI research as a demanding benchmark for algorithmic thinking, because the problems require novel solution approaches rather than pure factual knowledge. After an OpenAI system had previously placed just behind human top performers in competitions like the AtCoder Heuristic Division 2025, the clear victory in the Algorithm Division 2026 marks a turning point: for the first time, an AI fully solved all problems of an elite field in this format, while even the strongest human participants failed on two problems. It remains to be seen how well this capability transfers to software engineering tasks outside controlled competition conditions – for example, with internet access, larger codebases, or ambiguous requirements.
