Skip to content

Terminal-Bench 4.0

The current Terminal-Bench release measures difficult computer work after recalibrating task resources, fixing unstable tasks, and removing tasks that no longer separate frontier systems.

18 models

Higher is better.

  1. 1Claude Sonnet 5.5Anthropic70.60
  2. 2Claude Opus 5.5Anthropic66.40
  3. 3Claude Mythos 5.1Anthropic60.90
  4. 4GPT-6 AstraOpenAI57.90
  5. 5Gemini 4 ArgonGoogle57.40
  6. 6Claude Fable 5.1Anthropic55.80
  7. 7Pareto 26.9Unbiased51.00
  8. 8Pareto 26.10 PreviewUnbiased50.80
  9. 9Ling 3.1 FlashInclusionAI46.03
  10. 10DeepSeek V4.1 Flash 1017DeepSeek42.30
  11. 11Claude Haiku 5.5Anthropic39.20
  12. 12MiniMax M3.1 FlashMiniMax38.10
  13. 13Grok 4.7xAI38.00
  14. 14MiMo-V2.6-ProXiaomi34.90
  15. 15Step 5 PreviewStepFun33.30
  16. 16MiMo-V2.6-FlashXiaomi28.80
  17. 17SWE-2Cognition27.30
  18. 18Gemini 3.8 FlashGoogle19.10
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Task completion rate across 5 trials per task
Tasks
66 professional computer-work tasks
Authors
Ryan Marten, Terminal-Bench contributors, and Harbor (2026)

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