Google's Gemini 2.5 Deep Think Achieves Gold Medal at ICPC, Solves Problem Unsolved by Human Teams
"An advanced version of Gemini 2.5 Deep Think has achieved gold-medal performance at one of the world’s most prestigious programming contests," stated "nic" on social media, announcing a significant milestone for Google DeepMind's AI model. This breakthrough occurred at the 2025 International Collegiate Programming Contest (ICPC) World Finals, held on September 4 in Baku, Azerbaijan, where the AI demonstrated a profound leap in abstract problem-solving. Gemini 2.5 Deep Think secured a gold-medal level performance by correctly solving 10 out of 12 complex algorithmic problems.
The International Collegiate Programming Contest (ICPC) is globally recognized as the oldest, largest, and most prestigious algorithmic programming competition for college-level students. It attracts participants from nearly 3,000 universities across over 100 countries. During the intense five-hour event, teams face a set of intricate problems, with only perfect solutions earning points and time significantly influencing final rankings. Out of 139 competing teams in the finals, only the top four were awarded gold medals.
Competing in a remote online environment under official ICPC rules, Gemini 2.5 Deep Think solved 10 of the 12 problems within the allotted five-hour time constraint. Notably, the AI model successfully tackled Problem C, a complex optimization task that no human university team in the contest managed to solve. Had it been ranked alongside human teams, Gemini 2.5 Deep Think's performance would have placed it 2nd overall, solving eight problems in just 45 minutes and two more within three hours.
This achievement builds upon Gemini 2.5 Deep Think's earlier gold-medal win at the International Mathematical Olympiad (IMO) just two months prior, underscoring its advanced capabilities in abstract reasoning. Google DeepMind views these breakthroughs as a significant step towards artificial general intelligence (AGI), showcasing the AI's ability to synthesize novel solutions and demonstrate genuine ingenuity. The model's success is attributed to advances across pretraining, post-training, novel reinforcement learning techniques, multi-step reasoning, and parallel thinking.
The performance highlights the immense potential for AI to act as a true problem-solving partner for programmers, with experts suggesting that combining the best AI and human solutions could lead to solving all problems. A lightweight version of Gemini 2.5 Deep Think is already accessible to Google AI Ultra subscribers within the Gemini app. Google anticipates that smarter AI coding assistants will help developers tackle increasingly complex engineering challenges, fostering unprecedented collaboration between humans and AI.