MCPMark
Agents · 2025-09-28
MCPMark stress-tests agents on 127 realistic tasks that require create, read, update and delete operations through Model Context Protocol servers for Notion, GitHub, Filesystem, PostgreSQL and Playwright, each checked by a programmatic verifier. Scores report task success rate; server versions and tool-output truncation are recorded in score notes.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.7-Max (source unspecified) | 60.8 |
| Qwen3.7-Plus (source unspecified) | 58.7 |
| GPT-5.2 | 57.5 |
| GLM-5.1 | 57.5 |
| DeepSeek-V4-Pro | 57.1 |
| Opus 4.6 | 56.7 |
| Kimi K2.6 | 55.9 |
| GPT-5 | 50.9 |
| Qwen3.6 Plus (2026-04-02) | 48.2 |
| Qwen3.5 397B A17B | 46.1 |
| Opus 4.5 | 42.3 |
| DeepSeek-V3.2 | 38.0 |
| Qwen3.6 35B A3B | 37.0 |
| Qwen3.5-27B | 36.3 |
| Qwen3 Max Thinking (2026-01-23) | 33.5 |
| Sonnet 4.5 | 33.3 |
| GLM-5 | 31.1 |
| Kimi K2.5 | 29.5 |
| Qwen3.5 35B A3B | 27.0 |
| MiniMax M2 | 24.4 |
| Kimi K2 Thinking | 20.4 |
| Gemma 4 31B IT | 18.1 |
| Gemma 4 26B A4B IT | 14.2 |