Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals
Google DeepMind ● Covered by 3 sources
Google's Gemini 2.5 Deep Think just scored gold-medal results at the world's toughest college coding contest, ICPC World Finals. It solved a problem none of the 139 human teams cracked, and would've placed 2nd overall against university champs.
Based on reporting by Google DeepMind — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Two months after Gemini 2.5 Deep Think grabbed gold at the International Mathematical Olympiad, Google DeepMind sent it into an even tougher arena: the ICPC World Finals, the oldest and most prestigious college-level programming contest on the planet. This year's event, held in Baku on September 4, pulled in teams from nearly 3,000 universities across 103 countries. Only the top four out of 139 competing teams walked away with gold. Gemini, competing under the same five-hour clock and starting ten minutes after everyone else, solved 10 of 12 problems and landed in that gold tier.
The pace was the striking part. Gemini blew through eight problems in 45 minutes, then spent the remaining time grinding out two more over the next three hours, pulling from a deep toolkit of algorithms and data structures along the way. Add up the clock time across all 10 correct solutions — 677 minutes total — and Gemini would have ranked second overall against the human university teams, right behind whichever squad actually took first place.
The real headline, though, is Problem C, a nasty optimization puzzle about routing liquid through a duct network to fill reservoirs as fast as possible, with an infinite number of possible duct configurations to search through. Not a single university team solved it. Gemini cracked it in under 30 minutes by inventing a workaround: assign each reservoir a
My take — AI-written commentary, not fact-checked reporting
I'll say the quiet part: DeepMind loves stacking these gold-medal headlines back-to-back, IMO then ICPC, and it works as PR precisely because most people can't tell the difference between a curated demo and a repeatable capability. The Problem C solve is genuinely cool, minimax insight and all, but I'd rather see this Deep Think variant available outside a lab benchmark before I call it a step toward AGI. Impressive reasoning, sure. Overclaimed framing, also sure.
Read more about this at: Google DeepMind