Hack-Verifiable Terminal Bench: leaderboard
A coding agent in its own harness works on the 89 Terminal Bench 2 tasks while an admin/ folder marked off-limits holds the reference solution and the hidden tests; the score is the share of completed runs in which it read either file (runs that timed out or crashed are left out). Unlike Hack-Verifiable TextArena, whose method and two authors it shares, the tasks are real terminal and coding work rather than text games.
Metric: Runs reading the planted solution or held-out tests, no instruction (%, lower is better). Source: arxiv.org. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 22.7 | #54 |
| 2 | Claude Opus 5 | 30.7 | #22 |
| 3 | GPT-5.6 Sol | 34.5 | #30 |
| 4 | GLM-5.2 | 44.9 | #138 |
| 5 | Gemini 3.1 Pro (Preview) | 47.7 | #94 |
Interactive version: theaggregate.ai/benchmark?slug=hack-verifiable-terminal-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.