Skip to content
Hbench
hbench

Terminal-Bench 2.0 on Claude Opus 4.6

89 tasks, 5 attempts, open network. Updated 25 Oct 2026.

Build a run for this board

Value

What a task cost each configuration. pilot used 92% fewer tokens than Terminus 2, and 31% more time per task.

$ per solved task

Lower is cheaper. Dollars per task divided by the pass rate.

Embed this chart

Harness$ per solved task
Terminus 22.0.0 · Reference$1.05$0.66 a task
wozcode2.1.86$1.10$0.75 a task
meta-harness1.1.0$2.69$2.06 a task
Not reported: pilot, maya-v2, Judy, droid, Copilot CLI

Tokens per task

Lower is leaner. Input plus output, cached input included.

Embed this chart

HarnessTokens per task
pilot25k×0.08
Terminus 22.0.0 · Reference311k×1.00
wozcode2.1.86941k×3.02
meta-harness1.1.02.3M×7.47
Not reported: maya-v2, Judy, droid, Copilot CLI

Agent time per task

Lower is faster. Agent start to finish; setup and tests left out.

Embed this chart

HarnessTime per task
droid6.4 min×0.58
wozcode2.1.867.8 min×0.71
maya-v21.5-official9.1 min×0.83
meta-harness1.1.09.4 min×0.86
Terminus 22.0.0 · Reference10.9 min×1.00
Copilot CLI11.3 min×1.04
Judy0.7.012.7 min×1.16
pilot14.4 min×1.31

Pass rate against tokens per task

Up and to the left is better. Pass-rate axis starts at 50%.

Embed this chart

50%60%70%80%90%100%01.0M2.0M3.0MTokens per task, input and outputMore tokens for a lower pass rate than pilotpilotversion not recorded · 25k · 82.0%meta-harness1.1.0 · 2.3M · 76.4%wozcode2.1.86 · 941k · 68.1%Terminus 22.0.0 · 311k · 62.9%
  • On the frontier: no other row has a higher pass rate for less tokens
  • Off the frontier
  • 95% interval of the pass rate
On the frontier, least tokens first: pilot, 25k tokens a task, 82.0%. Not drawn, no tokens reported: maya-v2, Judy, droid, Copilot CLI.
  • 8 runs
  • Updated 25 Oct 2026
  • Method