sudo L7 (October 2026 rescoring)
Code
Surge's sudo L7 board as rescored between its October 23 and October 27, 2026 site builds, on 60 tasks in real codebases graded against expert-written rubrics. The board dropped every row's agent-harness tag, so Surge no longer says which harness ran each model, and these rows use configs with no harness. The page has no changelog and names no metric, pass threshold, grader or attempt count. The launch post quotes about 45% for Claude Opus 5.5 and Sonnet 5.5, 32% for GPT-6 Astra and 30% for GPT-6.1 Sol, which match neither this board nor the previous one. Every value fits 180 outcomes, which would be three attempts per task, though Surge states no count. Kept apart from the original board because the harness is no longer named and scores moved by as much as 8.9 points (Claude Sonnet 5.5 Max, 41.7 to 50.6; GLM 5.3 Max fell from 14.4 to 9.4).
Benchmark source: Surgehq