Skip to content
Hbench
hbench

Terminus 2

terminus-2

Terminus 2's pass rate on every board it ran on. Scores on different boards don't add up, so this page doesn't add them.

On each board

Terminal-Bench 2.0

Every board here: 2.0.0, full, 89 tasks, 5 attempts, open
Claude Opus 4.6
Ranks 6–8
8 harnesses62.9%54.4–70.8Terminus 2 2.0.0 is the reference.
deepseek-v3.2
39.6%31.5–48.1

HumanEvalFix

humanevalfix@1.0
Qwen2.5-Coder 7Bversion not recorded, sample-40, 40 tasks, 1 attempt, allowlist
7.5%0.0–17.5

How Terminus 2 works

Each hexagon shows how Terminus 2 works, ranked among the harnesses on that board: outward is better.

Terminal-Bench 2.0 on Claude Opus 4.6 (5 attempts, open network)

62.9%pass rate, 95% interval 54.4–70.8

Terminus 2 2.0.0, Lean contextTerminus 2 2.0.0, Cache useTerminus 2 2.0.0, Tests before finishingTerminus 2 2.0.0, Direct pathTerminus 2 2.0.0, No loopsTerminus 2 2.0.0, Speed

Compare with another harness

Terminus 2 2.0.0, Lean contextTerminus 2 2.0.0, Cache useTerminus 2 2.0.0, Tests before finishingTerminus 2 2.0.0, Direct pathTerminus 2 2.0.0, No loopsTerminus 2 2.0.0, Speedmeta-harness 1.1.0, Lean contextmeta-harness 1.1.0, Cache usemeta-harness 1.1.0, Tests before finishingmeta-harness 1.1.0, Direct pathmeta-harness 1.1.0, No loopsmeta-harness 1.1.0, Speed
  • Terminus 2 2.0.0 · 62.9%
  • meta-harness 1.1.0 · 76.4%