Unipat AI

A continuously updated benchmark evaluating AI coding agents on real-world software engineering tasks from GitHub issues.

State-of-the-art MLLMs achieve PhD-level language reasoning but struggle with visual tasks that 3-year-olds solve effortlessly.