MetroLLM-Bench: leaderboard
MetroLLM-Bench composite score on 238 held-out transit kiosk cases across six metro systems, where the model plans routes, quotes fares and handles disruptions through tool calls under operator-written policy.
Metric: Composite score (238 held-out cases, higher is better). Source: continker.ai. Saturation forecast: Around April 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Muse Glimmer 30B | 92.03 | #191 |
| 2 | Qwen 3.8 27B | 91.83 | #123 |
| 3 | Qwen 3.6 27B | 91.28 | #182 |
| 4 | Qwen 3.5 27B | 90.6 | #192 |
| 5 | GPT-5.6 Luna (Medium) | 90.57 | #79 (GPT-5.6 Luna) |
| 6 | GPT-5.4 (xHigh) | 90.45 | #83 (GPT-5.4) |
| 7 | Qwen 3.5 35B A3B | 89.9 | #268 |
| 8 | GPT-5.6 Sol (xHigh) | 89.82 | #22 (GPT-5.6 Sol) |
| 9 | GPT-5.4 (High) | 88.2 | #83 (GPT-5.4) |
| 10 | Qwen 3.5 9B | 88.05 | #391 |
| 11 | Mistral Small 4 | 87.82 | #522 |
| 12 | GPT-5.4 (Medium) | 87.72 | #83 (GPT-5.4) |
| 13 | Qwen 3.5 4B | 87.25 | #514 |
| 14 | GLM-4.7 Flash | 86.73 | #543 |
| 15 | GPT-5.4 Nano | 86.58 | #359 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=metrollm-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.