Terminal-Bench 2.1

What does it measure?
Whether an AI agent can finish a job inside a terminal (command line) with no human help: 89 tasks across 16 areas such as software engineering, security, scientific computing, data processing and debugging. It is the current evaluation Artificial Analysis uses after moving the older TerminalBench Hard to legacy.
Who evaluates it, and how?
Terminal-Bench 2.1 from Stanford and the Laude Institute. Each task runs in an isolated container where the AI works only by issuing shell commands; afterwards a hidden test suite in a separate container grades it, and the task counts only if every test passes. Maintainers keep fixing tasks found to be broken. The values here are Artificial Analysis's own runs.
Reading the score
0–100%. Top models now reach around 90%, so differences at the very top are small. Scores depend on the agent harness used with the model, so check that the run conditions match before comparing with Terminal-Bench numbers elsewhere.

Rank

Top score91.4%
Models tested81
Scoring basisHigher is better
Last updated2026.11.21
Model release date

Can it operate a computer itself to finish development jobs on its own?

Terminal-Bench 2.1

  1. 🥇AnthropicClaude Fable 5.191.4%
  2. 🥈OpenAIGPT-6 Astra89.9%
  3. 🥉OpenAIGPT-5.6 Sol89.5%
  4. #4AnthropicClaude Opus 589.1%
  5. #5SpaceXSIGrok 4.688.4%
  6. #6OpenAIGPT-5.6 Terra88.0%
  7. #7GoogleGemini 3.8 Flash87.6%
  8. #8AlibabaQwen3.8 Flash86.1%
  9. #9GoogleGemini 3.7 Flash85.8%
  10. #10MetaMuse Spark 1.385.4%
  11. #11Moonshot AIKimi K385.0%
  12. #12AnthropicClaude Fable 584.6%
  13. #12AnthropicClaude Opus 4.884.6%
  14. #14Z.aiGLM 5.3 Flash84.3%
  15. #14OpenAIGPT-5.584.3%
  16. #16Z.aiGLM 5.383.9%
  17. #17AnthropicClaude Opus 4.783.1%
  18. #18Sakana AIFugu Ultra82.1%
  19. #19SpaceXSIGrok 4.581.6%
  20. #20OpenAIGPT-5.6 Luna80.9%
  21. #21AnthropicClaude Sonnet 580.5%
  22. #22MetaMuse Spark 1.280.1%
  23. #23DeepSeekDeepSeek V4 Flash78.7%
  24. #23DeepSeekDeepSeek V4 Pro78.7%
  25. #23GoogleGemini 3.5 Flash78.7%
  26. #26OpenAIGPT-5.478.3%
  27. #27Z.aiGLM 5.277.9%
  28. #27MetaMuse Spark 1.177.9%
  29. #29GoogleGemini 3.6 Flash77.5%
  30. #30Motif TechnologiesMotif 374.9%
  31. #31AlibabaQwen3.7 Max74.5%
  32. #32GoogleGemini 3.1 Pro73.8%
  33. #33AnthropicClaude Sonnet 4.671.2%
  34. #34Moonshot AIKimi K2.7 Code67.4%
  35. #35Moonshot AIKimi K2.665.9%
  36. #36XiaomiMiMo V2.5 Pro65.2%
  37. #36MiniMaxMiniMax M365.2%
  38. #38TencentHy364.4%
  39. #39XiaomiMiMo V2.563.7%
  40. #40MetaMuse Spark62.2%
  41. #41Z.aiGLM-5.161.8%
  42. #42AlibabaQwen3.6 Plus61.4%
  43. #43AlibabaQwen3.7 Plus61.0%
  44. #44OpenAIGPT-5.4 Nano60.7%
  45. #45OpenAIGPT-5.4 Mini59.2%
  46. #46UpstageSolar Pro 457.3%
  47. #47AnthropicClaude Sonnet 4.555.8%
  48. #48MiniMaxMiniMax M2.755.4%
  49. #49NVIDIANemotron 3 Ultra53.9%
  50. #50GoogleGemini 3.5 Flash-Lite53.6%
  51. #51AlibabaQwen3.5 397B A17B51.3%
  52. #52Mistral AIMistral Medium 3.550.6%
  53. #53MeituanLongCat 2.050.2%
  54. #54DeepSeekDeepSeek V3.246.8%
  55. #55Moonshot AIKimi K2.545.7%
  56. #56AlibabaQwen3.6 35B A3B44.9%
  57. #57AnthropicClaude Haiku 4.544.2%
  58. #58GoogleGemma 4 31B43.4%
  59. #59InclusionAIRing-2.6-1T43.1%
  60. #60LG AI ResearchK-EXAONE 2.040.4%
  61. #61SpaceXSIGrok 4.339.7%
  62. #62StepfunStep 3.7 Flash39.3%
  63. #63SK TelecomA.X K239.0%
  64. #64NVIDIANemotron 3 Super38.6%
  65. #65AnthropicClaude Sonnet 436.3%
  66. #66CohereNorth Mini Code35.6%
  67. #67OpenAIGPT-535.2%
  68. #68GoogleGemini 3.1 Flash Lite31.1%
  69. #69LG AI ResearchK-EXAONE30.3%
  70. #70GoogleGemini 2.5 Pro28.5%
  71. #71OpenAIGPT OSS 120B26.2%
  72. #72Mistral AIMistral Small 421.0%
  73. #73Arcee AITrinity Large Thinking20.6%
  74. #74AmazonNova 2 Lite16.1%
  75. #75UpstageSolar Pro 312.0%
  76. #76MetaLlama 4 Maverick7.9%
  77. #77NVIDIANemotron 3 Nano 30B A3B6.7%
  78. #77NVIDIANemotron 3 Nano Omni6.7%
  79. #79OpenAIGPT-5 Mini3.7%
  80. #79MetaLlama 4 Scout3.7%
  81. #81IBMGranite 4.1 8B3.4%
Half of models ≤ 61.8%

59 of the 59 AIs released in the last 60 days have no score here yet · Gemini 4 Neon, GPT-6.2 Astra, Microsoft-Decision-1.1 and 56 more

This score comes from a fixed, predefined evaluation, so it may differ from what you get with your own question.

Scores labelled 'Self-reported' were published by the AI maker itself, not measured by a third party.

Related news