Coding and terminal agents
Terminal-Bench 2.1
89 difficult terminal tasks across engineering, systems and scientific work.
About Terminal-Bench 2.1
What it measures
Verifier-defined task success.
Scope & limitations
Keep 2.1 as a historical comparison track; it is not the latest release. Do not mix its results with 2.0.
Evaluation environment
Containerized command-line agent, Harbor and fixed resource budgets.