How long did it take the AI to solve all the problems?

Problems A, B, and C were solved relatively quickly, similar to the system's typical performance on earlier competitions where it solved everything in under an hour. Problems D and E, however, were exceptionally difficult; the system took about three hours to solve problem D, and finally cracked problem E after about three hours as well.

How does this year's result compare to last year's AtCoder competition?

About a year ago, an OpenAI AI placed second at the AtCoder Heuristics World Finals 2025 after running for ten hours fully autonomously. It lost its early lead midway through and was eventually overtaken by FakePsyho. This year, the system took first place with a wide margin.

What model powers this AI system?

The system consists of a model comparable to GPT-5.6 (which ships this Thursday) paired with a small harness to scale compute at test time. The system had no internet access. According to Borys Minaiev, six months ago they would not have been able to solve most of these problems.

AI Business & IndustryTHE DECODERJul 9, 2026Jul 9, 2026, 19:00 JST

OpenAI AI wins AtCoder competition, solving all five problems ahead of human finalists

Send on LINE

3 Key Points

What happened
An OpenAI AI system won the AtCoder World Tour Finals 2026 Algorithm Division exhibition match, solving all five problems and placing first with 8,300 points. Runner-up tour1st scored 4,300 points, and no human competitor solved problems C or E. Two problems (D and E) were rated exceptionally difficult and took the system about three hours to crack, though problems A, B, and C were solved more quickly.
Why it matters
This marks a major milestone in competitive programming — the system's performance continues a rapid climb. At the International Olympiad in Informatics (IOI) 2025, an OpenAI system ranked sixth among 330 human participants at gold medal level, climbing from the 49th percentile in 2024 to the 98th percentile a year later. At the 2025 ICPC World Finals, the system solved all twelve problems and would have taken first place. These results suggest AI reasoning capabilities are advancing significantly in domains requiring complex logical thinking.
What to watch
The system uses a model comparable to GPT-5.6, which ships this Thursday, paired with a small harness to scale compute at test time. OpenAI's next target is the International Olympiad in Informatics 2026 in early August. The system had no internet access and, according to Borys Minaiev (an ICPC world champion who works on reasoning models at OpenAI), would not have solved most of these problems six months ago.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

OpenAI's victory at AtCoder 2026 represents a striking continuation of a trend that has unfolded over the past year. The company's reasoning systems have moved from competitive but not dominant performances — finishing second in 2025 and narrowly missing bronze-level status at IOI 2024 — to gold-medal and first-place finishes across multiple prestigious competitions. The jump from the 49th percentile at IOI 2024 to the 98th percentile at IOI 2025 underscores the acceleration: in a single year, the system's rank among 330 participants jumped from roughly 162nd place to roughly 10th place.

What distinguishes this progression is that the improvements come without competition-specific training. At the ICPC, OpenAI stressed that the setup was an ensemble of general reasoning models, with GPT-5 solving eleven of the twelve problems and an experimental model handling the hardest problem after nine submissions — something no human team managed. This suggests the gains reflect underlying improvements in the reasoning capabilities of the base models themselves, not engineering optimized for a particular contest.

The specific difficulty of this year's AtCoder problems — particularly problems D and E, which took the system three hours despite solving A, B, and C more quickly — indicates the bar continues to rise. Borys Minaiev's comment that these two problems were significantly harder than any AtCoder problem the team had previously seen, combined with his observation that the system would not have solved most of them six months ago, points to a capability ceiling that is moving upward at a measurable pace. The next public test will be the International Olympiad in Informatics 2026 in early August.

FAQ

How long did it take the AI to solve all the problems?: Problems A, B, and C were solved relatively quickly, similar to the system's typical performance on earlier competitions where it solved everything in under an hour. Problems D and E, however, were exceptionally difficult; the system took about three hours to solve problem D, and finally cracked problem E after about three hours as well.
How does this year's result compare to last year's AtCoder competition?: About a year ago, an OpenAI AI placed second at the AtCoder Heuristics World Finals 2025 after running for ten hours fully autonomously. It lost its early lead midway through and was eventually overtaken by FakePsyho. This year, the system took first place with a wide margin.
What model powers this AI system?: The system consists of a model comparable to GPT-5.6 (which ships this Thursday) paired with a small harness to scale compute at test time. The system had no internet access. According to Borys Minaiev, six months ago they would not have been able to solve most of these problems.

THE DECODERRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.