What does a CursorBench score mean?

A CursorBench score is reported per reasoning-effort setting, not one number per model: the same model can post an identical score at two efforts for very different cost. On the live CursorBench 4.0 leaderboard, Opus 5.5 scores 56.0% at both High and Extra High, but Extra High costs $6.98 a task against High’s $3.97, nearly double for no accuracy gained. The highest-scoring row, usually a model’s Max setting, can hide a cheaper setting matching it.

The same pattern shows up in harness choice: what a model runs through changes its score as much as the model itself does, and reasoning effort is one more setting in that category, not a fixed property of the model being graded.

A single CursorBench percentage is a point on a cost curve, not a verdict on the model.