OpenAI’s Reasoning Model Achieves Gold-Medal Score at 2025 International Olympiad in Informatics

Key facts

• OpenAI’s reasoning model scored high enough to win a gold medal at IOI 2025, ranking first among AI participants.
• The 37th IOI was held in Sucre, Bolivia, from July 27 to August 3, 2025.
• The model finished sixth overall among 330 human contestants, trailing only five humans.
• OpenAI adhered to the same 5-hour time limit and 50-submission cap as human competitors.
• The model had no internet access, no RAG, and no IOI-specific training; it used a general reasoning model with only a basic terminal tool.
• In 2024, OpenAI’s o1-ioi model scored 213 points (49th percentile) under strict rules; in 2025, the model reached the 98th percentile.
• OpenAI’s model also achieved gold-level performance at IMO 2025 weeks earlier.
• OpenAI president Greg Brockman praised the model’s ‘gold-medal-level performance’ at IOI.

A Leap from Near-Bronze to Gold in One Year

OpenAI’s reasoning model has achieved a gold-medal score at the 2025 International Olympiad in Informatics (IOI), the premier global programming competition for secondary school students. Competing in the online AI track, the model placed sixth overall among 330 human contestants and first among all AI entrants. The 37th IOI took place in Sucre, Bolivia, opening on July 27 and closing on August 3, 2025. The Chinese national team swept the gold medals, but OpenAI’s model outperformed all but five human participants worldwide. The result marks a dramatic improvement from the previous year, when OpenAI’s model narrowly missed a bronze medal and ranked in the 49th percentile.

Strict Rules, No Special Treatment

OpenAI’s model competed under the same constraints as human participants: a five-hour time limit and a maximum of 50 submissions per problem. It had no access to the internet or retrieval-augmented generation (RAG), and was limited to a basic terminal interface. Unlike many AI systems that rely on extensive task-specific fine-tuning, OpenAI used a general-purpose reasoning model without any IOI-specific training. The only human intervention was selecting which solutions to submit and establishing a connection to the IOI API. This approach contrasts with last year’s strategy, which involved a more handcrafted test-time pipeline.

From 49th to 98th Percentile: The Numbers Behind the Jump

In 2024, OpenAI’s o1-ioi model scored 213 points under strict rules, placing in the 49th percentile. This year, the model’s score catapulted to the 98th percentile, a leap that underscores the rapid pace of improvement in AI reasoning. OpenAI’s research scientist Noam Brown revealed that the same model which achieved gold at the International Mathematical Olympiad (IMO) 2025 also proved to be the best competition programming model. “We did a comprehensive evaluation of the IMO gold model and found it performed best not only in contest math but also in many other areas, including programming,” Brown said.

The Technology Behind the Triumph

OpenAI’s reasoning models—o1, o1-ioi, and o3—were evaluated in a research paper titled “Competitive Programming with Large Reasoning Models,” now available on arXiv. The study explores how reinforcement learning (RL) can enhance complex coding and reasoning tasks without relying on human-designed test-time strategies. In CodeForce benchmarks, o1 achieved a score of 1673 (89th percentile), significantly outperforming non-reasoning models like GPT-4o and earlier reasoning models like o1-preview. The research found that increasing both RL compute and test-time reasoning compute consistently improved performance. Notably, o3 reached near-top-human level under strict submission limits without any manual strategy intervention, demonstrating that end-to-end RL can autonomously learn effective problem-solving techniques such as writing brute-force code for efficiency and cross-validating with alternative methods.

A String of Victories Across Competitions

The IOI gold-level performance follows OpenAI’s gold medal at IMO 2025 just weeks earlier. The company also achieved strong results at the AtCoder World Finals, marking a series of successes across multiple reasoning domains. OpenAI president Greg Brockman celebrated the model’s “gold-medal-level performance” at IOI, while the company stated it is working to build smarter, more capable models and bring them into mainstream products as quickly as possible. OpenAI’s rapid progress in competitive programming and mathematics signals a broader advance in AI reasoning capabilities, with implications for fields that require multi-step logic, self-play, and multi-agent collaboration.

Outlook: What the Achievement Means for AI and Education

The IOI 2025 result demonstrates that general-purpose reasoning models can now compete at the highest levels of human expertise in programming contests without domain-specific tuning. This raises questions about the future of competitive programming as a benchmark for human intelligence and about how educational systems might adapt. OpenAI has not announced a timeline for releasing the model to the public, but the company’s stated goal is to integrate these advances into its product lineup. The research paper emphasizes that RL-driven reasoning models can surpass handcrafted strategies, suggesting that further scaling of compute and training could yield even more dramatic improvements. For now, the gold-medal score at IOI stands as a milestone in the race toward artificial general intelligence.

The bottom line

• OpenAI’s general reasoning model scored gold-medal level at IOI 2025, ranking sixth among 330 human contestants and first among AI participants.
• The model improved from the 49th percentile in 2024 to the 98th percentile in 2025 without IOI-specific training or internet access.
• The same model that won gold at IMO 2025 also excelled at programming, showing that RL-trained reasoning models generalize across domains.
• OpenAI’s research paper on arXiv shows that increasing RL compute and test-time reasoning compute consistently boosts performance.
• The achievement signals that general-purpose AI can now compete with top human programmers under strict competition rules.