Terminal-Bench 2.0 · BenchLM

A benchmark for agentic software engineering tasks executed in real terminal environments. Models must inspect files, run commands, edit code, and recover from errors over multi-step workflows.

In index
Agentic

Models41

Top result82.0%GPT-5.5

Top-3 spread6.9 pts

YearN/A

Result and price

Results

41 results

#ModelResult

01

GPT-5.582.0%

02

GPT-5.3 Codex77.3%

03

GPT-5.475.1%

04

Qwen3.7 Plus70.3%

05

Qwen3.7 Max69.7%

06

Claude Opus 4.7 (Adaptive)69.4%

07CUComposer 2.569.3%

08

MiMo-V2.5-Pro68.4%

09

DeepSeek V4 Pro 081367.9%

10MAKimi K2.666.7%

11

MiMo-V2.565.8%

12

Claude Opus 4.665.4%

12

Qwen 3.6 Max (preview)65.4%

14

GLM-5.163.5%

15CUComposer 261.7%

16

Qwen3.6 Plus61.6%

17

GPT-5.4 mini60.0%

18

Qwen3.6-27B59.3%

18

Claude Opus 4.559.3%

20

Claude Sonnet 4.659.1%

21

Muse Spark59.0%

22

MiniMax M2.757.0%

23

DeepSeek V4 Flash 073156.9%

24

GLM-556.2%

25TEHy3 Preview54.4%

26

Qwen3.5 397B52.5%

27

Qwen3.6-35B-A3B51.5%

28MAKimi K2.5 (Reasoning)50.8%

28MAKimi K2.550.8%

30

Claude Sonnet 4.550.0%

31

Qwen3.5-122B-A10B49.4%

32

Grok 4.2047.1%

33

GPT-5.4 nano46.3%

34

MAI-Thinking-146.0%

35POLaguna M.145.8%

36

Qwen3.5-27B41.6%

37

GLM-4.741.0%

38

Qwen3.5-35B-A3B40.5%

39POLaguna XS 2.137.5%

40POLaguna XS.235.7%

41MELongCat-Flash-Lite-Sparse33.7%