GrandCode: AI System Beats All Humans in Competitive Programming

A new multi-agent reinforcement learning (RL) system called GrandCode has achieved a historic milestone in competitive programming by consistently outperforming all human participants in live contests. Developed by researchers and detailed in a paper on arXiv (2604.02721), GrandCode is the first AI system to achieve this feat. The system's success is attributed to two key innovations: orchestrating a variety of agentic modules—such as hypothesis proposal, solver, test generator, and summarization—and jointly improving them through post-training and online test-time RL. Additionally, the researchers introduced Agentic GRPO, a novel RL algorithm designed to handle multi-stage agent rollouts with delayed rewards and severe off-policy drift. This development marks a significant advancement in AI's coding capabilities, as competitive programming was previously considered one of the last human strongholds against AI. The best prior AI system, Google's Gemini 3 Deep Think, had achieved 8th place but was not evaluated under live competition conditions. GrandCode's achievement underscores the rapid progress in AI-driven problem-solving and raises questions about the future of human-AI competition in intellectual domains.