← All benchmarks

Long-Horizon Terminal-Bench 2.0 (mean reward)

Code · 2026-10-09

Long-Horizon Terminal-Bench (LHTB) 2.0 is 77 terminal tasks: the 46 LHTB 1.0 tasks plus 31 new ones, 30 of which ask the agent to reproduce a real scientific or engineering tool's behaviour from its observed outputs, in domains such as geospatial work, document and OCR pipelines, life sciences and physical simulation. An agent gets one attempt per task under a 3-hour budget in Terminus 2 on the maintainers' patched Harbor: a hidden verifier checks the work every 30 minutes and, until the task passes or time runs out, returns only a binary rejection and lets the agent continue. Original tasks keep their partial-credit graders and reconstruction tasks are graded on a six-stage ladder; the score is the mean reward over the 77 tasks, unscored tasks counting as 0, in percent. The maintainers state that its scores are not comparable with LHTB 1.0.

Benchmark source: GitHub

Top models (higher is better)