LoopArena: leaderboard
LoopArena (Alibaba DreamX, 2026) evaluates a language model as the controller of a fixed coding agent: after each round it reads a summary of the run and decides what the agent should do or verify next, or stops. This board is the strict success rate in percent on 27 full coding tasks from SCBench and BeyondSWE, three runs each, with Qwen3.7-Plus as the coding worker; higher is better.
Metric: Strict success rate (%; 27 full coding tasks from SCBench and BeyondSWE run from their original state (Type III); the evaluated model is the Controller: after each round it reads a Reporter's evidence packet and issues the next instruction to, or stops, a fixed Qwen3.7-Plus coding Worker, which alone has coding tools; a run succeeds only when it passes the task evaluator (every Core check for SCBench, reward 1 from the official Harbor evaluator for BeyondSWE) and follows the control protocol; three runs per task; provider-default reasoning). Source: arxiv.org. Saturation forecast: Around April 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 24.69 | #49 |
| 2 | Qwen 3.7 Plus | 23.46 | #207 |
| 3 | Claude Opus 4.8 | 20.99 | #56 |
| 4 | DeepSeek V4 Flash (0731) | 19.75 | #139 |
| 5 | GLM-5.2 | 16.05 | #138 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=looparena · How It Works · Data refreshed daily, snapshot 2026-11-13.