Gemini Deep Think reaches gold-medal level at ICPC World Finals
Working within the same five-hour limit given to student teams, the model would have placed second overall against the university competitors; OpenAI separately claimed a perfect score.
- Benchmarks & progress
- Models & capabilities
- Notable
Google DeepMind said Gemini 2.5 Deep Think solved 10 of 12 problems at the ICPC World Finals, within the same five-hour time limit given to the competing university teams, a result the company said matched gold-medal-level human performance — only the top four student teams earned gold. Had it been ranked among the human competitors, DeepMind said, the model’s combined solving time of 677 minutes would have placed it second overall.
One of the ten, Problem C, required optimising liquid flow through a network of interconnected ducts using dynamic programming and nested ternary search; DeepMind said no university team solved it. This followed Gemini’s gold-medal result at the International Mathematical Olympiad two months earlier, extending the same model line’s competition-mathematics performance into competitive programming.
OpenAI reported a separate, higher result at the same event: a combination of GPT-5 and an experimental reasoning model solved all 12 problems, also under the standard five-hour, PDF-based conditions given to students, with the hardest problem needing nine submission attempts before it was accepted by the official judge. The two companies’ results were not independently adjudicated against each other, and DeepMind’s post did not address OpenAI’s claim, leaving the comparison based on each lab’s own account of its performance under nominally the same contest rules.