DeployBench: leaderboard
DeployBench (2026) tests LLM agents in the OpenHands scaffold on setting up research artifacts released with published papers on a fresh Linux machine, verified by a hidden pipeline that runs the paper-designated experiment and checks its outputs. The board is the task success rate in percent on all 51 tasks; higher is better.
Metric: Task success rate (%) on all 51 research-artifact deployment tasks; the agent runs in the OpenHands scaffold on a fresh Linux VM (Docker disallowed) within a per-task time budget of 10 to 60 minutes, one run per task; a task passes when the hidden verifier runs the paper-designated experiment on the final VM state and checks its outputs; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.3 Codex | 51 | #116 |
| 2 | DeepSeek V4 Pro | 47.1 | #188 |
| 3 | Gemini 3.1 Pro (Preview) | 27.5 | #94 |
| 4 | Grok 4.20 | 11.8 | #297 |
| 5 | GPT-5.4 Mini | 7.8 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=deploybench · How It Works · Data refreshed daily, snapshot 2026-11-13.