Terminal-Bench 2.1 at 91.0% with Opus 5.5: how the number was produced
81 of 89 tasks with Opus 5.5 in a single sweep, refusals counted as fails, next to Claude Code (78) and Codex (75) on the same model, tasks and machine. Plus the earlier runs, the rerun rule, cost accounting, and why none of it is a leaderboard entry.