According to Zhidongxi on September 18, early this morning, OpenAI and Google announced that their models achieved gold medal-level performance in the finals of the world-renowned programming competition ICPC 2025 (49th International Collegiate Programming Contest).
The OpenAI reasoning system answered all 12 questions correctly, and answered 11 questions correctly at one time. It succeeded in solving the most difficult question after submitting it 9 times, ranking first compared with the human team ; the advanced version of Gemini 2.5 Deep Think solved 10 questions in 677 minutes, ranking second compared with the human team .
If AI is included in the overall ranking of ICPC, the top three should be OpenAI reasoning system, St. Petersburg State University, and Google Gemini 2.5 Deep Think advanced version .
ICPC requires contestants to solve 12 complex algorithm problems within 5 hours. The perfection of the solution and the time taken to solve the problem will affect the points.
In the end, the top four teams out of 139 won gold medals, namely Saint Petersburg State University, University of Tokyo, Beijing Jiaotong University, and Tsinghua University . Saint Petersburg State University solved the most questions, a total of 11.
The human team that won the ICPC gold medal
This is the second time that OpenAI's reasoning system and Google's Gemini 2.5 Deep Think have proven their strength in a top international competition, following their victory in the International Mathematical Olympiad (IMO) two months ago.
The code for Google's Gemini 2.5 Deep Think advanced version participating in the ICPC finals has been open sourced on GitHub.
GitHub address:
https://github.com/google-deepmind/gemini_icpc2025
01.
OpenAI scores full marks
Google got two things wrong
ICPC is recognized worldwide as the oldest, largest, and most prestigious university-level algorithm programming competition. Every year, participants from nearly 3,000 universities and more than 103 countries compete to solve real-world programming problems.
Both OpenAI and Google participated and achieved gold-level performance. OpenAI's reasoning system answered 12 questions, Google's Gemini 2.5 Deep Think advanced version answered 10 questions, and the best human team answered 11 questions .
1. OpenAI: Achieved a perfect score, answering all 11 questions correctly on the first try
OpenAI's reasoning system received a perfect score.
OpenAI mentioned that it did not train a model specifically for ICPC, and it used a combination of general reasoning models to participate in the competition.
During the competition, GPT-5 and an experimental reasoning model worked together to generate solutions, with the experimental reasoning model responsible for selecting the submitted solutions. Ultimately, GPT-5 correctly answered 11 questions, with the final and most difficult one being solved by the experimental reasoning model.
Its model answered 11 questions correctly in one go , with the most difficult question being successfully answered on the ninth submission.
2. Google: Answered 10 questions correctly, solved 8 questions in 45 minutes
Gemini 2.5, an advanced version of Deep Think, competed live in a remote online environment according to ICPC rules, starting 10 minutes after the human competitors. Gemini solved 10 of the 12 problems in 677 minutes, with eight of those taking 45 minutes and two taking three hours.
The figure below shows the time taken to solve each problem in the 2025 ICPC finals, with the Gemini time shown in blue and the fastest university team time shown in gray.
Gemini surpassed humans in solving all three problems .
The time it took to solve each problem in the ICPC finals
In addition, Google DeepMind also mentioned a difficult problem that had stumped all human teams, which was successfully solved by Gemini in half an hour.
Problem C requires teams to design a solution for delivering liquid to a set of tanks through a network of interconnected pipes. The goal is to find a pipe configuration that fills all the tanks as quickly as possible.
There are an infinite number of possible configurations for this problem, as each pipe can be open, closed, or even partially open, making it extremely difficult to find the optimal configuration.
Introduction to Problem C
Gemini has found an effective solution: it first assumes that each reservoir has a "priority value" that represents the degree of priority each reservoir should receive compared to other reservoirs.
When given a set of priority values, a dynamic programming algorithm can be used to find the optimal configuration of the pipeline.
Gemini discovered that by applying the Minimax Theorem, the original problem can be transformed into finding the priority value that can maximize the constraint on the final flow.
By leveraging the correlation between priority values and optimal flow, Gemini quickly found the optimal priority value through nested ternary searches in the bowl-shaped convex solution space, ultimately solving problem C.
Gemini users who currently subscribe to Google AI Ultra can already use the lightweight version of Gemini 2.5 Deep Think in the Gemini App.
02.
ICPC Gold Medal Level
Demonstrate abstract reasoning capabilities of large models
Google DeepMind's blog mentioned that Gemini's performance benefited from its technological innovations in pre-training, post-training, reinforcement learning technology, multi-step reasoning and parallel thinking .
For example, using reinforcement learning, researchers trained Gemini to reason and generate code for some of the most difficult problems faced by programmers, learning from the resulting feedback and improving its approach. To solve a problem, multiple Gemini agents would each propose their own solution, execute the code and test it using a terminal, and then iterate on the solution based on all the attempts.
Google DeepMind's internal research shows that the advanced version of Gemini 2.5 Deep Think can also achieve gold medal-level performance in the 2023 and 2024 ICPC World Finals, performing no less than the top 20 competitive developers in the world .
Achieving a gold medal-level result at ICPC has direct practical implications for software development. Combining the best AI and human solutions from the competition resulted in all 12 problems being solved thoroughly and correctly. This demonstrates the potential of AI to provide unique insights that complement human experts.
In addition to mathematics and programming, the advanced version of Gemini 2.5 Deep Think also demonstrates capabilities in abstract reasoning .
That’s because ICPC problems require models to understand complex problems, design multi-step logical plans, and execute them flawlessly—the same skills required in many fields of science and engineering, including designing new drugs or microchips.
OpenAI researchers posted on X that they used the same set of models to participate in the IMO and IOI competitions, demonstrating the model's performance and versatility.
03.
Conclusion: Large Models Are Complex
Improve abstract problem-solving skills
From the International Mathematical Olympiad (IMO) to this programming competition, OpenAI and Google's models have shown tremendous potential in solving increasingly challenging mathematical and reasoning problems. Dr. Bill Poucher, ICPC's Global Executive Director, stated that ICPC has always been committed to setting the highest standards in problem solving, and Gemini's achievements in this area mark a critical moment in defining the next generation of AI tools and academic standards.
Together, these breakthroughs in competitive programming and mathematical reasoning demonstrate that large models can achieve a performance leap in solving abstract reasoning problems, and may be able to collaborate with human experts to solve even more complex challenges.
This article comes from the WeChat public account "Zhidongxi" (ID: zhidxcom) , author: Cheng Qian, editor: Li Shuiqing, and is authorized to be published by 36Kr.
