Terminal Agentic Trajectory Generation on TerminalBench 2.0 benchmark leaderboard · SOTA2 Research

Terminal Agentic Trajectory Generation on TerminalBench 2.0

57.8ScoreClaude-Opus-4.5

Evaluation Results

Method	Links
Claude-Opus-4.52026.02		57.8
Gemini-3-Pro-Preview2026.02		56.9
Gemini-3.0 Pro2026.01		54.2
GPT-5.22026.02		54
Gemini-3-Flash-Preview2026.02		51.7
GPT-5.12026.02		47.6
DeepSeek-V3.22026.02		46.4
DeepSeek-V3.2 Thinking2026.01		46.4
Claude-Sonnet-4.52026.02		42.8
Claude Sonnet 4.52026.01		42.8
GLM-4.72026.02		41
MiMo-V2-Flash2026.01		38.5
Kimi-K2 Thinking2026.01		35.7
GPT-5 High2026.01		35.2
DeepSeek-V3.1-Nex-N12026.02		31.8
Minimax-M2.12026.02		29.2
Kimi-K2-Instruct2026.02		27.8
Qwen3-Coder-480B-A35B-Instruct2026.02		23.9
TerminalTraj-32B2026.02		22
TerminalTraj-14B2026.02		19.1
Qwen3-32B-Nex-N12026.02		16.7
Qwen3-Coder-30B-A3B-Instruct2026.02		14.6
Qwen3-235B-A22B-Instruct2026.02		13.5
TerminalTraj-7B2026.02		10.1
Qwen3-30B-A3B-Nex-N12026.02		8.3
Qwen2.5-Coder-32B-Instruct2026.02		4.49
Qwen3-32B2026.02		3.4
Qwen2.5-Coder-14B-Instruct2026.02		1.18
Qwen2.5-Coder-7B-Instruct2026.02		0