← All benchmarks

Long-Horizon Terminal-Bench 2.0 (solved, reward ≥ 0.95)

Code · 2026-10-09

The share of the 77 Long-Horizon Terminal-Bench 2.0 tasks an agent solves, counting a task when its final reward is at least 0.95 out of 1, in percent. On the reconstruction tasks every such reward is a full pass. It comes from the same single 3-hour Terminus 2 attempt per task as the mean-reward headline; the maintainers advise comparing solve counts rather than mean reward across the original and new task sets.

Benchmark source: GitHub

Top models (higher is better)