Skip to content
Hbench
hbench

Terminal-Bench 2.0 on GPT-5.3-Codex

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

Build a run for this board

Value

What a task cost each configuration.

$ per solved task

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

Embed this chart

Harness$ per solved task
spoox-o-m$3.42$2.44 a task
Not reported: droid, mux, Indusagi Coding Agent

Tokens per task

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

Embed this chart

HarnessTokens per task
mux472k
spoox-o-m1.2M
Not reported: droid, Indusagi Coding Agent

Agent time per task

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

Embed this chart

HarnessTime per task
droidReference7.4 min×1.00
mux8.3 min×1.12
Indusagi Coding Agent8.8 min×1.18
spoox-o-m14.6 min×1.97

Pass rate against tokens per task

Up and to the left is better. Not offered: dollars, which fewer than two rows report. Pass-rate axis starts at 60%.

Embed this chart

60%70%80%90%100%250k500k750k1.0M1.3M1.5MTokens per task, input and outputMore tokens for a lower pass rate than muxmuxversion not recorded · 472k · 74.6%spoox-o-mversion not recorded · 1.2M · 71.5%
  • 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: mux, 472k tokens a task, 74.6%. Not drawn, no tokens reported: droid, Indusagi Coding Agent.
  • 4 runs
  • Updated 25 Oct 2026
  • Method