Gemini 2.5 Deep Think hits gold medal level at ICPC coding finals

Two months after securing a gold medal at the International Mathematical Olympiad, Google DeepMind just sent its top-tier reasoning model into the fiercest competitive coding arena on Earth.

At the 2025 ICPC World Finals in Baku, Azerbaijan, an advanced version of Gemini 2.5 Deep Think went head-to-head against 139 elite university teams. Playing under official contest rules, the model started 10 minutes late, tackled the grueling five-hour problem set, and solved 10 out of 12 problems. With a total time of 677 minutes, it would have placed second overall, comfortably inside the four gold-medal slots.

A 45-minute blitz and a solo solve

The speed was absurd: Gemini knocked out eight problems in the first 45 minutes and picked up two more within three hours.

Even wilder was Problem C. The challenge involved optimizing liquid distribution through an interconnected network of ducts where valves could be fully open, closed, or partially open, creating an infinite search space. Exactly zero human university teams managed to solve it. Gemini cracked it in the first 30 minutes.

Its solution bypassed the infinite search space using an unexpected theoretical pivot: it assigned priority values to reservoirs, framed the bottleneck via the minimax theorem, and ran nested ternary searches across a convex solution space paired with dynamic programming.

Under the hood, DeepMind didn't just let a single model guess. It deployed a multi-agent system where multiple Gemini instances proposed candidate algorithms, executed code and tests directly inside isolated terminals, and iterated on their own errors in parallel. DeepMind reported that this same architecture also hits gold-medal marks on the 2023 and 2024 ICPC problem sets.

Why it matters

For programmers, AI assistants are officially graduating from glorified autocomplete to genuine problem-solving collaborators. While everyday users on Google AI Ultra can only access a lighter version of Gemini 2.5 Deep Think, the trajectory is clear: future developer tools will be capable of untangling gnarly algorithmic bottlenecks that stump senior engineers.

For the AI landscape, this moves the goalposts for reasoning benchmarks. Unlike standard code evaluation benchmarks where LLMs regurgitate memorized patterns, ICPC problems require novel mathematical synthesis under extreme time pressure and zero tolerance for edge-case errors. Gemini’s performance proves that combining reinforcement learning with multi-agent test-time compute can unlock abstract creativity once thought exclusive to human prodigies.

We think this puts immense pressure on rivals like OpenAI and Anthropic to step up their test-time execution environments. The caveat? Real-world software engineering isn't a five-hour clean algorithmic sprint; it's a messy marathon of fuzzy specs and legacy spaghetti code. But when it comes to raw computational logic, AI is no longer just studying human champions—it's beating them.