OpenAI's Reasoning System Achieves Gold at 2025 International Olympiad in Informatics (IOI)

OpenAI has made a significant leap forward in AI capabilities, with its reasoning system achieving gold medal status in the 2025 International Olympiad in Informatics (IOI). This marks a notable improvement from last year's performance and highlights the rapid advancements in general-purpose reasoning models.

Technical Breakdown

Competition Overview

• Event: Online AI track of the IOI 2025
• Participants: 330 human contestants, several AI systems
• Performance: OpenAI's system placed first among AI participants and ranked higher than all but five human contestants.
• Constraints:
  • 5-hour time limit
  • 50 submission limit
  • No internet access or Retrieval-Augmented Generation (RAG)
  • Access to a basic terminal tool

Model Architecture and Strategy

• Ensemble of Models: OpenAI used an ensemble of general-purpose reasoning models. These models were not specifically trained for the IOI, showcasing their versatility.
• Scaffolding:
  • Selection of solutions to submit
  • Connection to the IOI API

Progress and Improvement

• 2024 Performance: Last year, OpenAI's system narrowly missed a bronze medal, placing in the 49th percentile.
• 2025 Performance: This year, the system achieved gold, placing in the 98th percentile. The significant improvement is attributed to advancements in research methods and more efficient test-time strategies.

Recent Successes

OpenAI has seen a string of successes in various competitive programming events:

• AtCoder World Finals
• International Mathematical Olympiad (IMO)
• IOI

These achievements demonstrate the effectiveness of OpenAI's efforts to build smarter, more capable models and integrate them into mainstream products.

Team Effort

The success is a result of collaborative work within OpenAI. Key contributors include:

• Preparation for IOI: @alexwei_, @bminaiev, @oleg_murk
• Long-term Work on Competitive Programming: @_lorenzkuhn, @MostafaRohani, @clavera_i, @andresnds, @ahelkky

Personal Reflections

Sheryl Hsu, a senior engineer at OpenAI, shared her excitement about the team's achievements:

• Personal Experience: Sheryl and some teammates traveled to Bolivia to attend the IOI in person. They had the opportunity to meet participants and coaches, making the experience even more rewarding.
• Contest Adherence: The system competed under the same conditions as human contestants, without direct supervision from contest organizers.

Conclusion

OpenAI's success at the 2025 IOI underscores the rapid progress in AI reasoning capabilities. By leveraging general-purpose models and refining test-time strategies, OpenAI has demonstrated significant improvements over just one year. This achievement is a testament to the team's dedication and the potential of AI in competitive programming.