DynamicMCPBench: leaderboard
DynamicMCPBench (ITMO University, 2026) scores LLM agents on realistic tasks over live Model Context Protocol servers by whether they reproduce the effects of a recorded successful trajectory, never the final answer text, under deterministic replay. Pass^3 in percent, a task counts only if all three attempts succeed; higher is better.
Metric: Pass^3 (%; share of the 750 balanced tasks from 15 tool-use categories over 121 MCP servers solved in all three independent attempts; effect checkpoints scored deterministically under replay, no LLM judge; tools needed plus eight distractors offered). Source: arxiv.org. Saturation forecast: Around June 2027. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.7 Max | 51.2 | #145 |
| 2 | GLM-5.1 | 50.3 | #179 |
| 3 | Qwen 3.6 35B A3B | 48.5 | #462 |
| 4 | DeepSeek V4 Pro | 46.4 | #188 |
| 5 | Gemma 4 31B | 42.5 | #394 |
| 6 | MiniMax-M3 | 42.4 | #269 |
| 7 | Claude Haiku 4.5 | 41.1 | #568 |
| 8 | Kimi K2.6 | 40.7 | #178 |
| 9 | Grok 4.3 | 31.6 | #333 |
| 10 | Qwen 3.5 4B | 27.3 | #1175 |
| 11 | GPT-5.4 Mini | 25.7 | #389 |
| 12 | Qwen 3 8B | 22.1 | #1830 |
| 13 | Gemma 4 E4B | 22.1 | #2214 |
| 14 | Qwen 2.5 7B | 21.1 | #2703 |
| 15 | Gemma 4 E2B | 20.1 | #3892 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dynamicmcpbench · How It Works · Data refreshed daily, snapshot 2026-11-13.