Coding benchmark

Terminal-Bench 4.0 leaderboard

16 models have a published score. Claude Sonnet 5.5 (Anthropic) leads with 70.6, and the spread to the lowest score, Gemini 3.8 Flash at 19.7, is 50.9 points. The best open-weight result, Kimi K3.1 Pro at 65.7, trails the best closed model, Claude Sonnet 5.5, by 4.9. The highest-scoring model you can deploy on OpenLLM Buddy is MiMo V2.6 Pro, ranked 13 of 16 with 34.9. Sources report different settings for this test, such as test year, version, or reasoning effort. Each setting is listed in the source table, so compare close scores with care.

Best openllm buddy catalog

34.9

MiMo V2.6 Pro

2 models scored

Best open weights

65.7

Kimi K3.1 Pro

5 models scored

Best closed api

70.6

Claude Sonnet 5.5

9 models scored

Terminal-Bench 4.0 scores

Percentage of tasks completed. Scores are far lower than on version 2.x and are kept on a separate leaderboard.

  • OpenLLM Buddy catalog
  • Open weights
  • Closed API
  1. 1Claude Sonnet 5.5Terminal-Bench 4.0; safeguards enabled70.6
  2. 2Claude Opus 5.5Terminal-Bench 4.0; xhigh effort (66.36% in text, safeguards enabled)66.4
  3. 3Kimi K3.1 ProTerminal-Bench 4.0, Claude Code harness, max effort65.7
  4. 4GPT-6 AstraTerminal-Bench 4.0; AA independent run (Max)59.1
  5. 5Gemini 4 ArgonTerminal-Bench 4.0 (AA-run figure as shown in Google's table)57.1
  6. 6GPT-6.1 SolTerminal-Bench 4.0; AA independent run (Max)56.1
  7. 7Claude Fable 5.1Terminal-Bench 4.0; max effort, Claude Code --bare, 15 trials/task (Mythos 5.1 scored 60.9%)55.8
  8. 8GLM 5.3Terminal-Bench 4, Max, mini-swe-agent harness41.9
  9. 9Mistral Large 4Terminal-Bench 4.0, ML4 OSS release (Preview scored 28)41.0
  10. 10Claude Haiku 5.5Terminal-Bench 4.0; safeguards enabled39.2
  11. 11MiniMax M3.1 FlashTerminal-Bench 4.0, Codex 0.149.1 harness38.1
  12. 12Grok 4.7Terminal-Bench 4.0; xhigh effort37.6
  13. 13MiMo V2.6 ProTerminal Bench 4.034.9
  14. 14Step 5 PreviewTerminal-Bench v4, High effort (vendor-reported)33.3
  15. 15DeepSeek V4.1 FlashTerminal-Bench 4.0, Pass@1, max effort31.2
  16. 16Gemini 3.8 FlashTerminal-Bench 4.0; AA independent run (High)19.7

What it measures

A refreshed set of long, multi-step terminal tasks built after models began clearing most of version 2. Tasks run in sandboxed containers and are checked by scripts, and the agent harness used is noted next to each score.

How it is scored

Percentage of tasks completed. Scores are far lower than on version 2.x and are kept on a separate leaderboard.

Why it matters

With 2.x scores bunched in the 80s, version 4.0 is where frontier and open agents are separating now.

Official benchmark page

More coding tests

Sources

ModelTestScoreSettingSource
Claude Sonnet 5.5Terminal-Bench 4.070.6Terminal-Bench 4.0; safeguards enabledAnthropic Claude Sonnet 5.5 System Card (sec 8.5); also in launch post
Claude Opus 5.5Terminal-Bench 4.066.4Terminal-Bench 4.0; xhigh effort (66.36% in text, safeguards enabled)Anthropic Claude Opus 5.5 System Card (Table 8.1.A / sec 8.5)
Kimi K3.1 ProTerminal-Bench 4.065.7Terminal-Bench 4.0, Claude Code harness, max effortMoonshot AI blog: Kimi K3.1 Pro
GPT-6 AstraTerminal-Bench 4.059.1Terminal-Bench 4.0; AA independent run (Max)Artificial Analysis
Gemini 4 ArgonTerminal-Bench 4.057.1Terminal-Bench 4.0 (AA-run figure as shown in Google's table)Google DeepMind Gemini models page (benchmark table)
GPT-6.1 SolTerminal-Bench 4.056.1Terminal-Bench 4.0; AA independent run (Max)Artificial Analysis
Claude Fable 5.1Terminal-Bench 4.055.8Terminal-Bench 4.0; max effort, Claude Code --bare, 15 trials/task (Mythos 5.1 scored 60.9%)Anthropic Claude Fable 5.1 / Mythos 5.1 System Card (sec 8.6)
GLM 5.3Terminal-Bench 4.041.9Terminal-Bench 4, Max, mini-swe-agent harnessHugging Face model card (GLM-5.3), evaluation results
Mistral Large 4Terminal-Bench 4.041.0Terminal-Bench 4.0, ML4 OSS release (Preview scored 28)Mistral AI blog: Mistral Large 4 open weights
Claude Haiku 5.5Terminal-Bench 4.039.2Terminal-Bench 4.0; safeguards enabledAnthropic Claude Haiku 5.5 System Card (sec 8.4); also in launch post
MiniMax M3.1 FlashTerminal-Bench 4.038.1Terminal-Bench 4.0, Codex 0.149.1 harnessHugging Face model card (MiniMax-M3.1-Flash)
Grok 4.7Terminal-Bench 4.037.6Terminal-Bench 4.0; xhigh effortxAI Introducing Grok 4.7 (benchmark table)
MiMo V2.6 ProTerminal-Bench 4.034.9Terminal Bench 4.0MiMo-V2.6-Pro-RL model card (Hugging Face)
Step 5 PreviewTerminal-Bench 4.033.3Terminal-Bench v4, High effort (vendor-reported)cellcog.ai (relaying StepFun's release table)
DeepSeek V4.1 FlashTerminal-Bench 4.031.2Terminal-Bench 4.0, Pass@1, max effortDeepSeek-V4.1-Flash model card (Hugging Face)
Gemini 3.8 FlashTerminal-Bench 4.019.7Terminal-Bench 4.0; AA independent run (High)Artificial Analysis

Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works