sudo L7
2026-10-09 · General coding ability
- What it measures
- General coding ability: making changes in real codebases, spotting risk, verifying dangerous cases, and explaining the call. 60 tasks written by staff-level engineers, mostly from private production repositories, plus open source. On this site it is a general board and appears in both the Web frontend and Android scenarios.
- How
- Surge AI's sudo L7. Experts write a rubric for each task. Models grade against it, with unit tests, static analysis, and other executable checks as evidence. The blog describes eight dimensions and names functional correctness, practical judgment, thought partnership, verification, engineering craft, truthful reporting, and communication.
- What the score means
- The percent is the share of assignments completed without missing essential requirements. The leaderboard page publishes only that column. Functional correctness for the top four is close: Sol 82.4%, Opus 5.5 82.1%, Astra 81.5%, Sonnet 5.5 79.6%. Thought partnership is 76.1% for Opus and 74.8% for Sonnet, against 53.0% for Sol and 51.6% for Astra. Communication is 72.4% for Sonnet and 44.1% for Sol.
- Caveats
- The board is new, the set is 60 tasks, and the tasks stay private. Models grade against expert rubrics. Surge is an AI data vendor. Per-task costs in the blog are approximate; the leaderboard page has no cost column.
- Updates
- The blog was published on 2026-10-08. The leaderboard page was last published on 2026-10-09. This page synced the 12 public rows on 2026-10-10.
Full board
12
| # | Model | Score |
|---|