All Benchmarks/Agents

MCPMark

Older versionHigher is better

The earlier edition of a test of whether the AI can complete 127 tasks using real MCP servers such as GitHub, Notion, and PostgreSQL. It has been superseded by the verified edition. Scores run from 0 to 100%. Higher is better.

A newer version is available: MCPMark-Verified

Other versions:MCPMark-Verified
Top score52.6%GPT-5
Models tested12
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    OpenAI
    GPT-5
    2025.08.07 · medium
    52.6%
  2. 2
    Anthropic
    Claude Opus 4.5
    2025.11.24 · high
    42.3%
  3. 3
    DeepSeek
    DeepSeek V3.2
    2025.12.01 · thinking
    36.8%
  4. 4
    Anthropic
    Claude Sonnet 4.5
    2025.09.29
    32.1%
  5. 5
    OpenAI
    GPT-5 Mini
    2025.08.07 · high
    30.3%
  6. 6
    Anthropic
    Claude Opus 4.1
    2025.08.05
    29.9%
  7. 7
    Anthropic
    Claude Sonnet 4
    2025.05.22 · high
    28.3%
  8. 8
    Google
    Gemini 2.5 Pro
    2025.06.17
    15.8%
  9. 9
    Google
    Gemini 2.5 Flash
    2025.06.17
    9.1%
  10. 10
    OpenAI
    GPT-4.1
    2025.04.14
    8.1%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.