CodeRouterBench - LLM Router (Online Agent): leaderboard
Model routing for coding tasks from Agent-as-a-Router (Zhou et al., 2026): the model reads each of the 2,919 in-distribution CodeRouterBench coding tasks and picks which of eight candidate LLMs should solve it, as an agent with a nearest-neighbour memory of past routing outcomes, warmed up on 200 validation tasks and updated online as the test stream proceeds. Score is the average recorded task score in percent of the chosen models, where a missing result counts 0; higher is better.
Metric: Average task score (%) of the backend the model routes each task to, over the 2,919 in-distribution CodeRouterBench coding tasks (execution-graded or proxy scores of eight candidate LLMs, a missing result counting 0); the router picks one of the eight by name; the zero-shot router with a cosine-kNN memory warmed up on 200 validation tasks and updated online as the test stream proceeds, seed 42). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniMax-M2.7 | 45.27 | #519 |
| 2 | GLM-5 | 45.21 | #256 |
| 3 | Qwen 3 Max | 44.92 | #401 |
| 4 | Qwen 3.5 Plus | 43.98 | #227 |
| 5 | Claude Sonnet 4.6 | 43.65 | #158 |
| 6 | Kimi K2.5 | 42.7 | #252 |
| 7 | GPT-5.4 | 40.95 | #121 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=coderouterbench-llm-router-online-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.