JobBench

Citation frequency 2/3
What does it measure?
Tests how well an AI agent (an AI that uses tools to carry out a task end to end) performs tasks close to real workplace work.
Who evaluates it, and how?
65 tasks drawn from 35 occupations, which the AI completes with tools. Grading criteria also inspect the output files. The maintainers run every model in the same OpenCode environment.
Reading the score
Scores run from 0 to 100%. Higher is better.
Top score75.4%
Models tested30
Scoring basisHigher is better
Last updated2026.09.22
Model release date

Can it do real work from many different jobs?

JobBench

  1. 🥇AnthropicClaude Opus 5.575.4%
  2. 🥈MetaMuse Wonder75.3%
  3. 🥉AnthropicClaude Fable 5.173.1%
  4. #4OpenAIGPT-6 Astra66.1%
  5. #5AnthropicClaude Opus 565.7%
  6. #6MetaMuse Spark 1.364.9%
  7. #7XiaomiMiMo-V2.6-Pro62.0%
  8. #8TencentHy4Reasoning effort: Max61.7%
  9. #9MetaMuse Spark 1.261.6%
  10. #10Z.aiGLM 5.361.4%
  11. #11XiaomiMiMo-V2.6-Flash61.2%
  12. #12StepfunStep-5Reasoning effort: High59.0%
  13. #13AnthropicClaude Fable 557.4%
  14. #14AlibabaQwen3.8 Flash55.7%
  15. #15MetaMuse Spark 1.154.7%
  16. #16Moonshot AIKimi K354.3%
  17. #17AnthropicClaude Opus 4.848.4%
  18. #18OpenAIGPT-5.6 Sol45.4%
  19. #19AnthropicClaude Opus 4.744.5%
  20. #20Z.aiGLM 5.243.4%
  21. #21OpenAIGPT-5.538.3%
  22. #22AnthropicClaude Sonnet 4.636.6%
  23. #23OpenAIGPT-5.432.2%
  24. #24GoogleGemini 3.5 Flash31.5%
  25. #25AlibabaQwen3.7 Plus27.6%
  26. #26NVIDIANemotron 3 Ultra27.2%
  27. #27XiaomiMiMo V2.5 Pro25.0%
  28. #28AnthropicClaude Sonnet 4.520.7%
  29. #29MetaMuse Spark17.0%
  30. #30GoogleGemini 3.1 Pro15.9%
Half of models ≤ 54.5%

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.