Gemini reaches gold-medal level in ICPC world finals

Gemini 2.5 Deep Think solved 10 of the 12 problems in the ICPC world finals and would have placed second among the 139 participating universities, according to Google DeepMind.

The model competed remotely on September 4, during the final held in Baku, Azerbaijan. It started 10 minutes after the human teams and had to work under the same rules and within the same five-hour limit.

What “gold-medal level” means

The ICPC is one of the most demanding university programming competitions. Every year, it brings together teams from nearly 3,000 universities in more than 103 countries.

Participants must solve algorithmic problems by writing programs that work correctly. There are no partial points: a solution only counts if it is accepted, and the time taken also affects the ranking.

Of the 139 teams in the 2025 final, only the top four received gold medals. Gemini solved eight problems in the first 45 minutes and two more over the following three hours. With a total time of 677 minutes, it would have placed second if compared directly with the university teams.

A problem no human team solved

The most notable performance was on problem C. None of the university teams managed to solve it, but Gemini found a solution during the first half-hour.

The challenge involved distributing liquid through a network of pipes to fill several tanks as quickly as possible. Each pipe could be open, closed or partially open, creating a virtually infinite number of possible configurations.

The model turned the problem into a more manageable search. It assigned each tank a priority value, calculated the best pipe configuration for those values and then used a combination of the minimax theorem, dynamic programming and ternary searches to approach the optimal solution.

In simple terms, it did not test every combination. It found a way to describe the problem that drastically reduced the search space.

What changes for people who program

This result points to an AI capable of doing more than completing lines of code or explaining a function. Gemini had to understand a new problem statement, design a strategy, write the program, run it, check it and correct it when necessary.

Google explains that the system used several agents that proposed different solutions, ran code and tests in terminals and reviewed the results before producing a final answer. It also combined reinforcement training, multistep reasoning and parallel exploration of alternatives.

The practical value is clear: a programming assistant could help detect difficult bugs, design algorithms for logistics or explore solutions in scientific research. That does not mean it can automatically replace a team of engineers. In this kind of competition, a single incomplete solution is worthless.

In fact, Google says that combining Gemini’s best solutions with those of the human teams would have correctly solved all 12 problems. The most reasonable interpretation is that the model can contribute ideas that complement experts, not that human results are no longer necessary.

A result announced by Google with an important caveat

DeepMind presents this achievement as progress in abstract reasoning and connects it to its path toward artificial general intelligence. It also says that a similar version would have reached gold-medal level in the 2023 and 2024 finals.

But there is an important nuance: the ICPC organization confirmed that the submitted solutions were complete and accepted, while clarifying that its review did not validate the system, the processes or the underlying model. The result therefore supports the submitted solutions, but it is not equivalent to an independent audit of the entire experiment.

For you, the important point is the direction this result indicates: programming assistants are beginning to tackle problems that require planning, verification and algorithmic creativity. The next step will be finding out how much of this capability holds up outside a competition, in real projects with changing requirements and errors that do not have a single answer.