Terminal-Bench 4.0 leaderboard

Every task has its own verifier and passes only if every check passes, so there is no partial credit; timeouts and unverifiable submissions score zero, and refusals count as failures. TensorFeed reports the Vals AI run, which uses one harness (mini-swe-agent) for every model and averages three runs, so the column compares like with like. With its September 29, 2026 update Vals moved the whole board from the Terminus 2 harness to mini-swe-agent, and every score here was replaced in one pass, so figures quoted before that date are not comparable. Two things move the number more than people expect. The harness: Claude Fable 5 went from 22.73 to 41.41 when Vals switched harnesses, and the official board, where each lab runs its own agent, differs again. And fallback: when a safeguard reroutes a task to another model, Vals counts the attempt, and the per-cell note says how many attempts that was. Terminal-Bench 2.x and 4.0 share no tasks, so never compare scores across versions.