An advanced version of Gemini 2.5 Deep Think has achieved gold-medal level performance at the 2025 International Collegiate Programming Contest (ICPC) World Finals, marking a major leap in AI’s ability to solve complex, abstract problems. Competing under the same rules and time constraints as human university teams, the AI system solved 10 out of 12 problems in just 677 minutes—finishing in second place overall if ranked against human teams. The contest, held in Baku, Azerbaijan, is the most prestigious algorithmic programming competition for college students, with 139 teams from 103 countries participating. Only four teams earned gold medals, making Gemini’s achievement particularly significant. One of the most remarkable moments came when Gemini solved Problem C—uniquely, the only problem no human team managed to crack. This optimization challenge involved finding the ideal configuration of open, closed, or partially open ducts to fill multiple reservoirs as quickly as possible. With an infinite number of possible configurations, the problem required deep insight and mathematical creativity. Gemini devised a novel solution by assigning priority values to reservoirs and applying the minimax theorem. It then used nested ternary searches within a convex solution space to efficiently identify optimal configurations—demonstrating a level of ingenuity previously unseen in AI. This success follows Gemini 2.5 Deep Think’s earlier gold-medal performance at the 2025 International Mathematical Olympiad (IMO), reinforcing its growing prowess in high-level reasoning. The AI’s performance in both competitions showcases a profound advancement in abstract thinking, creativity, and multi-step problem solving—key traits associated with artificial general intelligence (AGI). The breakthrough stems from a suite of innovations in pretraining, post-training, reinforcement learning, and parallel reasoning. Gemini employs multiple AI agents that generate, test, and refine code iteratively. Each agent runs code in isolated environments, evaluates results, and shares feedback to improve subsequent attempts. This process, combined with advanced algorithmic strategies and dynamic programming, enables the system to explore complex solution spaces effectively. Google DeepMind emphasized that this achievement is not just a technical milestone but a step toward redefining academic and competitive standards. Dr. Bill Poucher, ICPC Global Executive Director, praised the result as a pivotal moment in shaping the future of AI tools and education, calling it a catalyst for a digital renaissance. While the AI solved 10 problems correctly, the competition’s strict rules—only perfect solutions earn points, and every minute counts—highlight the high bar it met. Gemini’s ability to solve eight problems within 45 minutes and two more within three hours underscores its speed and precision. A lightweight version of Gemini 2.5 Deep Think is already available to Google AI Ultra subscribers via the Gemini app, offering early access to these advanced capabilities. The ICPC has confirmed the validity of the submitted solutions, though it did not verify the underlying system or model. This achievement builds on foundational research across Google DeepMind’s Gemini, Thinking, and Coding teams, involving hundreds of engineers and researchers. The project was led by Hanzhao Lin and Heng-Tze Cheng, with key contributions from experts in competitive programming, machine learning, and system architecture. The results demonstrate that AI is no longer just mimicking human problem solving—it is outperforming it in certain domains, particularly those requiring deep abstraction and innovation. As these capabilities evolve, they promise to empower students, researchers, and developers with powerful tools for discovery and creation. This milestone signals a turning point in AI’s role in education, science, and engineering, bringing the vision of AGI closer to reality.