At the 2025 International Collegiate Programming Contest (ICPC) World Finals, OpenAI and Google DeepMind demonstrated just how far large language models (LLMs) have advanced. Competing in ICPC’s official AI track, their systems solved complex algorithmic challenges that pushed beyond the limits of human performance.

Both the AI teams and the human teams had five hours to complete the challenge, with rankings determined by the number of problems solved and total time taken, including penalties for incorrect submissions. Every team received the same set of 12 algorithmic problems, which were judged under identical conditions through the ICPC online system.

The ICPC World Finals drew teams from 139 universities across 103 countries. Contestants were given five hours to solve a uniform set of twelve algorithmic problems, with rankings based on both accuracy and speed.

Human gold medal winners, all solving their problems within the same five-hour limit, included:

The highest-ranked U.S. teams — Harvard University and MIT — placed at the silver medal level.

OpenAI’s GPT-5 completed all 12 problems within the 5-hour limit, marking the first time in ICPC history that any competitor — human or AI — achieved a perfect score. The feat was the equivalent of winning a gold medal. GPT-5 solved 11 of the 12 problems on its own, while the final, most difficult challenge was completed with assistance from an experimental reasoning model after multiple submissions. OpenAI emphasized that GPT-5 was not specifically trained on ICPC problems, underscoring its strength as a general-purpose model. No human team managed the same.

Google’s Gemini 2.5 Deep Think — entered in an advanced version — solved 10 problems in 677 minutes (11 hours and 17 minutes), a performance Google said would have ranked second overall. Gemini solved eight problems within just 45 minutes and added two more within three hours, before stalling on the final two. Most notably, Gemini solved “Problem C,” which no human team completed. The challenge involved distributing liquid through a series of ducts, requiring a creative mathematical approach. Google explained that Gemini applied the minimax theorem and nested ternary searches to discover an optimal solution, highlighting its ability to devise strategies beyond human attempts.

While LLMs such as GPT-5 and Gemini are already known for excelling at benchmarks and general knowledge, the ICPC results mark a notable shift into advanced, unsolved algorithmic challenges.

Earlier in 2025, Google reported that Gemini won a gold medal at the International Mathematical Olympiad, underscoring rapid progress in mathematical reasoning. Yet only months earlier, leading models had struggled on the FrontierMath benchmark, failing to solve higher-level problems. The ICPC results suggest that the gap between AI models and humans is closing quickly.

For enterprises, the ICPC performance demonstrates the potential of LLMs in automating increasingly complex workflows. While most businesses don’t need models that can solve the world’s hardest programming challenges, the same reasoning and problem-solving skills can be applied to tasks such as optimization, advanced data analysis, and strategic decision-making. Having models that have proven strong coding and mathematical capabilities makes AI systems more useful, adaptable, and trustworthy for real-world applications.

Many observers see these achievements as a step toward artificial general intelligence (AGI), where models display reasoning capabilities comparable to humans.

The ICPC results suggest LLMs are narrowing that gap. As OpenAI and Google continue pushing the boundaries, these systems are gaining attention across social media and enterprise communities alike.

Q: What is the ICPC?
A: The International Collegiate Programming Contest is the world’s largest coding competition, bringing together top university teams to solve algorithmic problems.

Q: How did GPT-5 perform?
A: OpenAI’s GPT-5 completed all 12/12 problems — something no human team has ever achieved. It solved 11 directly, and an experimental reasoning model finished the last one after multiple submissions.

Q: How did Gemini 2.5 perform?
A: Google’s Gemini 2.5 Deep Think — entered in an advanced version — solved 10/12 problems in 677 minutes (11 hours and 17 minutes), a performance that would have ranked second overall. Unlike GPT-5, which relied primarily on its standard public model, Gemini required its advanced version to reach this result.

Q: Did the models compete directly with humans?
A: No. They competed in a separate AI track but faced the same problems, rules, and time constraints.

Q: Why does this matter for enterprises?
A: Success at the ICPC suggests that LLMs can handle complex, unsolved algorithmic challenges, pointing to potential in advanced enterprise applications such as workflow automation, optimization, and analysis.

The ICPC results mark more than a symbolic victory — they show that AI systems can now match or surpass humans in domains that demand deep reasoning, persistence, and creativity. For enterprises, this could translate into AI tools capable of handling mission-critical workflows such as supply chain optimization, financial modeling, and engineering design — problems that are closer in spirit to unsolved algorithmic challenges than to routine automation.

Equally significant is the precedent: no human team in ICPC history has ever solved all 12 problems, yet GPT-5 did. Achievements like this reinforce the idea that foundation models are not just statistical engines but evolving systems capable of discovering novel strategies.

For researchers and business leaders alike, the contest results highlight a narrowing gap between AI and human intelligence. While artificial general intelligence remains a debated concept, the trajectory is clear — models are moving steadily closer to capabilities once thought uniquely human.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.