Terminal-Bench v3.0Coding

Official public leaderboard; each model uses the highest-scoring thinking level reported by the Terminal Bench 3.0 authors. Gemini results are self-computed with a mini swe agent (LiteLLM 1.96).Terminal-Bench 3.0 is the third-generation terminal-agent benchmark in the Terminal-Bench series. Its tasks cover general agent capabilities and evaluate how models autonomously complete complex tasks in real terminal environments. Each model's score uses the highest-scoring thinking level reported by the benchmark authors.

Ranking Sources

#ModelScore

1

GPT-5.6 SolOpenAI

34.6%

2

Claude Fable 5Anthropic

34.1%

3

GLM-5.3Zhipu AI

28.3%

4

Grok 4.6xAI

26.0%

5

Claude Opus 4.8Anthropic

21.1%

6

GPT-5.6 TerraOpenAI

20.8%

7

Kimi K3Moonshot AI

17.4%

8

Grok 4.5xAI

15.7%

9

Gemini 3.7 FlashGoogle

14.9%

10

Claude Sonnet 5Anthropic

14.6%

12 models total

Back to benchmarks

Terminal-Bench v3.0Coding

Official website

Official public leaderboard; each model uses the highest-scoring thinking level reported by the Terminal Bench 3.0 authors. Gemini results are self-computed with a mini swe agent (LiteLLM 1.96).Terminal-Bench 3.0 is the third-generation terminal-agent benchmark in the Terminal-Bench series. Its tasks cover general agent capabilities and evaluate how models autonomously complete complex tasks in real terminal environments. Each model's score uses the highest-scoring thinking level reported by the benchmark authors.

Ranking Sources

#ModelScore

1

GPT-5.6 SolOpenAI

34.6%

2

Claude Fable 5Anthropic

34.1%

3

GLM-5.3Zhipu AI

28.3%

4