Inference
EN

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.

ModelsModels with a result.41
Top resultBest result on this benchmark.82.0%GPT-5.5
Top-3 spreadPoints from first to third.6.9 pts
YearYear of release.—

Result and price

36 of 41 results

Results

41 results
#ModelResultThe score from the source.
01GPT-5.582.0%
02GPT-5.3 Codex77.3%
03GPT-5.475.1%
04Qwen3.7 Plus70.3%
05Qwen3.7 Max69.7%
06Claude Opus 4.7 (Adaptive)69.4%
07Composer 2.569.3%
08MiMo-V2.5-Pro68.4%
09DeepSeek V4 Pro 081367.9%
10Kimi K2.666.7%
11MiMo-V2.565.8%
12Claude Opus 4.665.4%
12Qwen 3.6 Max (preview)65.4%
14GLM-5.163.5%
15Composer 261.7%
16Qwen3.6 Plus61.6%
17GPT-5.4 mini60.0%
18Qwen3.6-27B59.3%
18Claude Opus 4.559.3%
20Claude Sonnet 4.659.1%
21Muse Spark59.0%
22MiniMax M2.757.0%
23DeepSeek V4 Flash 073156.9%
24GLM-556.2%
25Hy3 Preview54.4%
26Qwen3.5 397B52.5%
27Qwen3.6-35B-A3B51.5%
28Kimi K2.5 (Reasoning)50.8%
28Kimi K2.550.8%
30Claude Sonnet 4.550.0%
31Qwen3.5-122B-A10B49.4%
32Grok 4.2047.1%
33GPT-5.4 nano46.3%
34MAI-Thinking-146.0%
35Laguna M.145.8%
36Qwen3.5-27B41.6%
37GLM-4.741.0%
38Qwen3.5-35B-A3B40.5%
39Laguna XS 2.137.5%
40Laguna XS.235.7%
41LongCat-Flash-Lite-Sparse33.7%