Harness-efficiency benchmark · Terminal-Bench 2.x

Can your harness ship it for less?

Everyone runs the same agent — a vanilla piCLI on OpenRouter. Tune the harness, not the model, to solve the ranked Terminal-Bench task set for the fewest billed dollars. Match the GLM 5.2 baseline's pass count, then cheapest wins— and every dollar is real, pulled from OpenRouter's generation ledger.

$0.9143cheapest qualifying · baseline
10tasks passed
50harnesses submitted
Cheapest qualifying runrecord so far: $0.9143 · $/task (low → high)
$0.82$1.04$1.27$1.49$1.71submissions →

The line is the record — the cheapest qualifying harness so far, stepping down as entrants get cheaper. Dots are submissions (greener = more tasks passed) — click one to open its run.

Leaderboard32 entries · cheapest qualifying harness per participant
#ParticipantPassedBilled cost$/passed
Below the bar18 entries · ran but under 10 passed
ParticipantPassedBilled costTo qualify

How it works

1

Same agent for everyone

Every entry runs a vanilla pi CLI on OpenRouter. The model isn't the variable — the harness is.

2

Tune the harness

Add skills, profiles, config — anything but the model source — to solve the ranked Terminal-Bench task set your way.

3

Cheapest wins

Match the GLM 5.2 baseline's pass count, then the fewest billed OpenRouter dollars takes the top spot.