The ICPC World Finals was dominated by AI. The GPT-5 combined system solved all 12 problems correctly and topped the rankings, while humans could only fight tooth and nail for the third place.
This generation of college students has it tough. After struggling to make it to the programming competition finals, they're outshined by AI.
At the just - concluded 2025 International Collegiate Programming Contest (ICPC) World Finals, OpenAI's system perfectly solved all 12 problems. If included in the rankings, it would rank first.
The Google Gemini 2.5 Deep Think model solved 10 problems, reaching the gold - medal standard and ranking second.
This top - level event brought together 139 top teams from nearly 3,000 universities in 103 countries around the world.
In the independent "AI experimental track" supervised by ICPC, the AI systems faced the same problems and evaluation criteria as human contestants and performed very impressively.
One of the more difficult problems, "Problem C", was not solved by any university team, but was solved by both the Gemini and OpenAI model combinations.
For some unknown reason, Google started its system 10 minutes later than the human contestants. That's a bit excessive.
OpenAI Wins with a Perfect Score, Pairing GPT - 5 with a Powerful External Aid
OpenAI sent a combined lineup of GPT - 5 and an experimental reasoning model this time. In less than five hours of competition time, this system successfully solved all 12 problems.
Specifically, GPT - 5 independently completed the first 11 problems, and 11 of them passed the ICPC official online evaluation environment on the first submission.
The most challenging problem for the OpenAI system was Problem G. After GPT - 5 encountered difficulties, the experimental reasoning model took over. The two models submitted a total of 9 times to solve this difficult problem.
Among human contestants, only the first - place team from St. Petersburg State University solved it in 3 attempts. Many university teams didn't even submit a solution for this problem once.
OpenAI hasn't shared the solution idea for this problem yet, but they specifically emphasized that the participating model is a general - purpose reasoning model and wasn't specially trained for ICPC.
It's also worth mentioning that many researchers in the OpenAI team are former ICPC contestants themselves, including Chief Scientist Jakub Pachocki, who succeeded Ilya. You can read about his story in "Altman: Praises Two Poles, OpenAI Hasn't Met a Problem They Can't Solve".
It's been almost exactly a year since OpenAI first launched the reasoning model o1. Compared with a year ago, today's reasoning models are not only more intelligent and faster but also more cost - effective.
Google Gemini Solves Difficult Problems Independently
Google Gemini 2.5 Deep Think started 10 minutes after the competition began and finally solved 10 out of 12 problems within five hours, reaching the gold - medal level.
Gemini solved 8 problems within the first 45 minutes and another 2 within three hours. Calculated based on a total time of 677 minutes, if compared with university teams, Gemini's result would rank second.
In the following figure, the time taken by Gemini to solve problems is shown in blue, and the time taken by the fastest university team is shown in gray.
For Problem C, which human contestants failed to solve, Google shared the solution idea.
This problem requires designing a plan to distribute liquid to storage repositories through an interconnected pipeline network, with the goal of finding a configuration that fills all storage repositories as quickly as possible.
Since each pipeline can be opened, closed, or partially opened, the number of possible configurations is infinite, making it extremely difficult to search for the optimal configuration.
Gemini's solution:
It first assumes that each storage repository has a "priority value", representing the degree to which it should be prioritized relative to other storage repositories.
Given a set of priority values, a dynamic programming algorithm can be used to find the optimal configuration of the pipelines. By applying the minimax theorem, the original problem can be solved by finding the priority values that make the resulting flow most constrained.
Using the relationship between priority values and optimal flow, Gemini uses nested ternary search to quickly find the optimal priority values in the bowl - shaped convex solution space, successfully solving Problem C.
Google DeepMind said that Gemini's success combines a series of technological advancements such as pre - training, post - training, new reinforcement learning techniques, multi - step reasoning, and parallel thinking.
During the reinforcement learning process, they trained Gemini to reason and generate code for some of the most difficult problems faced in the programming world, learning from the feedback of the results and continuously improving the methods.
One More Thing
Since the advent of the reasoning model paradigm, AI has continuously performed well in the International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the programming competition ICPC.
However, Research Vice - President Jerry Tworek of OpenAI said this time:
After ICPC, we probably won't participate in other competitions. The next frontier is more exciting.
Human contestants can breathe a sigh of relief.
So what is the next frontier? You can refer to Tworek's latest personal introduction:
How to apply various scientific and engineering skills to real - world problems.
Researchers and workers in these fields, here comes GPT - 5.
Reference Links:
[1]https://x.com/OpenAI/status/1968368133024231902
[2]https://deepmind.google/discover/blog/gemini-achieves-gold-level-performance-at-the-international-collegiate-programming-contest-world-finals/
[3]https://worldfinals.icpc.global/2025/
This article is from the WeChat public account "QbitAI", author: Meng Chen. Republished by 36Kr with permission.