RideWay: leaderboard
RideWay (2026) is a Chinese-language ridehailing benchmark that scores tool-using agents on interaction efficiency beyond task success, over 58 stateful tool-agent-user tasks gated by a three-judge LLM success panel. Efficiency Utility is a success-gated score in which a failed trajectory earns zero and a successful one is discounted for excess tool calls and user-facing turns beyond human reference floors, with the two penalties calibrated from blinded human pairwise preferences and reported here on the held-out split; higher is better.
Metric: Efficiency Utility (success-gated, held-out custom14 mean over 14 tasks x 3 attempts; a successful trajectory is discounted for excess tool calls and user-facing turns beyond human reference floors, penalties fit from human pairwise preferences; failures score 0). Source: arxiv.org. Saturation forecast: Around February 2027. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Non-reasoning) | 70.5 | #48 (GPT-5.5) |
| 2 | Qwen 3.7 Max (Non-reasoning) | 66.1 | #142 (Qwen 3.7 Max) |
| 3 | Qwen 3.6 27B (Non-reasoning) | 63.8 | #339 (Qwen 3.6 27B) |
| 4 | Claude Sonnet 5 | 63.7 | #108 |
| 5 | Gemma 4 31B | 62.8 | #393 |
| 6 | Qwen 3.7 Plus (Non-reasoning) | 62.4 | #203 (Qwen 3.7 Plus) |
| 7 | GLM-5.2 (Non-reasoning) | 62 | #137 (GLM-5.2) |
| 8 | Qwen 3.6 Plus (Non-reasoning) | 61.2 | #217 (Qwen 3.6 Plus) |
| 9 | Claude Opus 4.6 | 60.6 | #100 |
| 10 | MiMo-V2.5-Pro (Non-reasoning) | 59 | #274 (MiMo-V2.5-Pro) |
| 11 | Kimi K2.6 (Non-reasoning) | 56.3 | #173 (Kimi K2.6) |
| 12 | DeepSeek V4 Pro (Non-reasoning) | 54.3 | #181 (DeepSeek V4 Pro) |
| 13 | Gemma 4 26B | 49.3 | #614 |
| 14 | Gemini 3.5 Flash (Minimal) | 44.3 | #98 (Gemini 3.5 Flash) |
| 15 | Qwen 3.5 9B (Non-reasoning) | 43.1 | #803 (Qwen 3.5 9B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rideway · How It Works · Data refreshed daily, snapshot 2026-11-12.