SPACEXAI
4/5 workers samplingPOSSIBLE_MODEL_CHANGE · gpt-5
public · agents

Terminal Bench

End-to-end terminal tasks in isolated sandboxes.

Items
80
Revision
2026.06
Contamination
watched
Scorer
scorer.term.v2
Dataset revision
[email protected]
Scorer
env-oracle · scorer.term.v2
License
Apache-2.0
Kind
Public reproduction suite
ModelScore$/taskMedian sToolsEvidence
Claude 4.1 Opus58.6$0.32850.43.8LIVE_MEASURED
Grok 4 Heavy53.1$0.15065.62.8LIVE_MEASURED
Claude 4.1 Sonnet52.4$0.05834.03.4LIVE_MEASURED
GPT-551.4$0.08137.63.0LIVE_MEASURED
Gemini 2.5 Pro49.2$0.02925.32.9LIVE_MEASURED
Grok 448.8$0.04830.03.2LIVE_MEASURED
o4-pro44.6$0.695107.02.2LIVE_MEASURED
Mistral Large 343.4$0.02229.02.9LIVE_MEASURED
DeepSeek V3.141.2$0.005123.42.7REPRODUCED
Qwen3-Max40.8$0.007027.72.8LIVE_MEASURED
Llama 4 405B38.2$0.007026.52.6REPRODUCED
GPT-5 Mini36.8$0.004414.72.4LIVE_MEASURED
Gemini 2.5 Flash34.6$0.00169.62.2LIVE_MEASURED
Grok 3 Mini31.2$0.002511.52.1LIVE_MEASURED
Llama 4 70B28.6$0.00165.42.0REPRODUCED