Overview

• The model correctly solved 10 of 12 World Finals problems, which would have placed it second overall against human teams.
• It ran live in a remote environment under ICPC rules with organizer oversight, executing and iterating code during the five-hour contest.
• Gemini uniquely solved Problem C by assigning priority values to reservoirs and combining dynamic programming, minimax reasoning, and nested ternary search to locate optimal flows.
• Google attributes the result to advances in reinforcement learning, multi-step reasoning, parallel solution generation, and automated code testing; the company published solutions for review on GitHub.
• The company describes the performance as progress toward AGI, while reporters note Google has not disclosed the compute or cost required for the extended evaluation.