Skip to content

AutomationBench

An agent benchmark for completing automation workflows in reproducible task environments.

25 models

Higher is better.

  1. 1MiniMax M3.1 FlashMiniMax62.2
  2. 2DeepSeek V4.1 Flash 1017DeepSeek60.5
  3. 3Atria Dawn PreviewShanghai Artificial Intelligence Laboratory53.8
  4. 4MiMo-V2.6-ProXiaomi53.1
  5. 5Ling 3.1 FlashInclusionAI52.5
  6. 6MiMo-V2.6-FlashXiaomi52.3
  7. 7Gemini 4 ArgonGoogle51.3
  8. 8Ultramarine 1 FlashPaper Instruments50.2
  9. 9Muse Spark 1.3Meta49.4
  10. 10GLM-5.3-FlashZ.AI48.8
  11. 11GLM-5.3Z.AI48.2
  12. 12Step 5 PreviewStepFun44.0
  13. 13GPT-6 AstraOpenAI41.4
  14. 14Claude Opus 5.5Anthropic40.0
  15. 15GPT-6.1 SolOpenAI36.1
  16. 16GPT-6 SolOpenAI33.2
  17. 17Hy4 previewTencent32.1
  18. 18DeepSeek V4 Pro 0813DeepSeek31.8
  19. 19Claude Fable 5.1Anthropic31.4
  20. 20Kimi K3Moonshot AI30.8
  21. 21Gemini 3.7 FlashGoogle30.4
  22. 22LFM2.5-24B-A2BLiquid AI27.7
  23. 23Qwen3.8 MaxAlibaba27.3
  24. 24Claude Opus 5Anthropic26.0
  25. 25DeepSeek V4 Flash 0731DeepSeek25.1
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Agent task-completion score
Tasks
600 public automation tasks
Authors
Moonshot AI (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.