// BENCHMARK

Terminal-Bench 2.0 benchmark

AI model leaderboard for the Terminal-Bench 2.0 benchmark. Compare how large language models score on Terminal-Bench 2.0, see the full ranking, and understand what this AI benchmark measures. GLM-5.2 currently leads with 81%. Long-horizon shell/agent benchmark testing multi-step terminal automation.

Leaderboard

# Model Organization Score Variant Source
#1 GLM-5.2 Zhipu AI 81%
#2 Claude Opus 4.7 Anthropic 68.5% official ↗
#3 Kimi K2.6 Moonshot 66.7% official ↗
#4 Qwen3.6-Max-Preview Qwen 63.2% official ↗
#5 Qwen3.6-Plus Qwen 61.6% official ↗
#6 Claude Opus 4.5 Anthropic 59.3% official ↗
#7 DeepSeek V4-Pro DeepSeek 58.1% official ↗
#8 MiniMax M2.7 MiniMax 57% official ↗
#9 GLM-5.1 Zhipu AI 55.3% official ↗
#10 Gemini 3 Pro Google DeepMind 54.2% cited-deepseek-v3-2 official ↗
#11 Gemini 3 Pro Google DeepMind 54.2% cited-devstral-readme official ↗
#12 MiniMax-M2.5 MiniMax 50.1% official ↗
#13 Kimi K2 Thinking Moonshot 47.1% thinking-w-simulated-tools official ↗
#14 DeepSeek-V3.2 DeepSeek 46.4% cited-devstral-readme official ↗
#15 DeepSeek-V3.2 DeepSeek 46.4% thinking official ↗
#16 MAI-Thinking-1 Microsoft 46% official ↗
#17 Kimi K2 Moonshot 44.5% k2-0905-w-simulated-tools-cited-k2-thinking official ↗
#18 Claude Opus 4 Anthropic 43.2% official ↗
#19 Claude Sonnet 4.5 Anthropic 42.8% cited-deepseek-v3-2 official ↗
#20 Claude Sonnet 4.5 Anthropic 42.8% cited-devstral-readme official ↗
#21 LongCat-Flash Meituan Inc 39.51% official ↗
#22 DeepSeek-V3.2-Exp DeepSeek 37.7% from-hf-readme official ↗
#23 GLM 4.5 Zhipu AI 37.5% official ↗
#24 North Mini Code 1.0 Cohere 36%
#25 Kimi K2 Thinking Moonshot 35.7% cited-devstral-readme official ↗
#26 Kimi K2 Thinking Moonshot 35.7% thinking-cited-deepseek-v3-2 official ↗
#27 Claude Sonnet 4 Anthropic 35.5% official ↗
#28 GPT-5 OpenAI 35.2% cited-deepseek-v3-2 official ↗
#29 Devstral 2 Mistral AI 32.6% from-hf-readme official ↗
#30 GPT-4.1 OpenAI 30.3% official ↗
#31 o3 OpenAI 30.2% official ↗
#32 GLM-4.5-Air Zhipu AI 30% official ↗
#33 MiniMax-M2 MiniMax 30% cited-devstral-readme official ↗
#34 MiniMax-M2 MiniMax 30% cited-deepseek-v3-2 official ↗
#35 gemini-2.5-pro Google DeepMind 25.3% official ↗
#36 Kimi K2 Moonshot 25% official ↗
#37 GLM 4.6 Zhipu AI 24.6% cited-devstral-readme official ↗
#38 Devstral Small 2 Mistral AI 22.5% from-hf-readme official ↗
#39 DeepSeek-R1-0528 DeepSeek 17.5% official ↗

Frequently asked questions about Terminal-Bench 2.0

What is the Terminal-Bench 2.0 benchmark?

Long-horizon shell/agent benchmark testing multi-step terminal automation.

How is the Terminal-Bench 2.0 benchmark scored?

Terminal-Bench 2.0 is scored using the success-rate (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on Terminal-Bench 2.0?

As of the latest reported scores on GenAIList, GLM-5.2 achieves the highest result on Terminal-Bench 2.0 with a score of 81%.

Is a higher Terminal-Bench 2.0 score better?

Yes. On Terminal-Bench 2.0 a higher score indicates better performance, so models near the top of the leaderboard are the strongest.