Terminal-Bench
Claude Opus 4.8
Anthropic
GPT-5.4
OpenAI
Gemini 3 Pro
Google DeepMind
Claude Sonnet 4.6
Anthropic
DeepSeek V4
DeepSeek
Qwen3 Max
Alibaba
Llama 4.1 405B
Meta
Kimi K2.5
Moonshot AI
Scores are illustrative sample data.
What it measures
Terminal-Bench measures whether a model can actually operate a machine: diagnose a broken environment, run the right command chain, and leave the system in a verifiably correct state. Tasks span system diagnosis, environment repair, filesystem verification, and concurrency.
Methodology
Pinned environments — Every task boots from the same initial system image, so runs stay deterministic and comparable across models.
Trace + evidence — The full command chain and resulting system state are kept, with log evidence for every step.
Objective verification — A scripted checker measures the end state against the target — no human grading, no credit for plausible-looking output.
← Back to Leaderboard