Benchmarks & Deep Dives · Free

Terminal-Bench

Terminal-Bench is an open-source benchmark for AI agents working in a command-line environment. Each task ships with its own Docker environment, a human-verified reference solution and tests that check the end state, covering work such as building software, configuring systems, data analysis and scientific computing. It is hosted by Stanford, Harbor and the Laude Institute, runs on the open-source Harbor evaluation framework, and is versioned as a continuously maintained benchmark (1.0 in May 2025, 2.0 in November 2025, 3.0 and 4.0 in 2026). Scores measure a model and agent harness together.