GPT-5 and Gemini 2.5 Ace ICPC World Finals, Outperforming Human Teams in Algorithmic Challenges

In a groundbreaking demonstration of AI’s capabilities, OpenAI’s GPT-5 and Google DeepMind’s Gemini 2.5 Deep Think participated in the 2025 International Collegiate Programming Contest (ICPC) World Finals. Both models not only outperformed human teams but also achieved scores that would have placed them at the top of the competition.

What Changed Technically

• GPT-5 and Gemini 2.5's Performance:
  • GPT-5 scored a perfect 12 out of 12, solving all problems within the five-hour time limit.
  • Gemini 2.5 solved 10 out of 12 problems in 677 minutes, which would rank it second overall if it were competing directly with human teams.

Why It Matters to Practitioners

• Algorithmic Problem Solving:
  • The ICPC is known for its complex and previously unsolved algorithmic challenges. The success of these LLMs in such a rigorous environment demonstrates their potential in real-world applications where efficient problem-solving is crucial.
  • This performance could signal a shift towards using LLMs for tasks that traditionally require human expertise, potentially freeing up developers to focus on higher-level strategic work.

Competition Details

• ICPC World Finals:

  • The competition attracts thousands of participants from around the globe. In 2025, 139 universities from at least 103 countries competed in the World Finals.
  • Teams are given a set of algorithmic problems to solve within a five-hour time frame. Final rankings are based on both the number of solved problems and the speed of completion.

• LLM Participation:

  • Both GPT-5 and Gemini 2.5 participated under ICPC rules, with their performance supervised by OpenAI and Google DeepMind.
  • The LLMs were provided with the same PDF problem set as human teams and submitted solutions to the ICPC World Finals Local Judge.

Implementation Notes

• GPT-5:

  • OpenAI did not train a specific version of GPT-5 for the ICPC. Instead, they used an existing model that had already been fine-tuned on a broad range of tasks.
  • The system’s first answer was correct for 11 out of the 12 problems. For the most challenging problem, it succeeded on its ninth submission.

• Gemini 2.5 Deep Think:

  • Google entered an “advanced version” of Gemini 2.5, which had been optimized for complex algorithmic tasks.
  • The model’s performance was particularly notable given the time constraints and the difficulty of the problems.

Human Performance

• Top Teams:
  • St. Petersburg State University, the University of Tokyo, Beijing Jiaotong University, and Tsinghua University were the gold medal winners among human teams.
  • Harvard and MIT secured silver medals, highlighting the high level of competition.

None of the human teams achieved a perfect score, with GPT-5’s performance standing out as a significant achievement in the field of AI.

Conclusion

The success of GPT-5 and Gemini 2.5 at the ICPC World Finals underscores the rapid advancements in large language models. These results suggest that LLMs are not only capable of understanding and generating human-like text but can also tackle complex, real-world problems with high efficiency and accuracy. As AI continues to evolve, we can expect to see more applications where these models complement or even surpass human capabilities.