← All benchmarks

Revelo Code Index

Code · 2026-09-25

Revelo's Code Index scores coding agents, each a model in a named agent harness, on 60 private tasks that Revelo's senior engineers wrote and reviewed in three 20-task sets following the SWE-Lancer, Terminal-Bench 2.1 and Terminal-Bench 3.0 formats. Every pairing runs each task three times in Harbor containers graded by the task's executable verifier; a trial passes when its verifier reward is at least 1, a set's pass rate is the mean over its tasks of the per-task trial mean, and the index is the unweighted mean of the three set pass rates, which equals the share of passing trials over all 60 tasks. This three-set definition dates from September 25, 2026; the July 29 launch index also averaged a SWE-Bench-Pro Hard set, which Revelo removed because agents had internet access.

Top models (higher is better)