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

Agent Episodes

Browser, repo, spreadsheet, API, and simulator episodes. human_io tracked.

Items
96
Revision
2026.08
Contamination
clean
Scorer
scorer.agent.v3
Dataset revision
[email protected]
Scorer
task-consequence · scorer.agent.v3
License
GBU public suite
Kind
Public reproduction suite
ModelScore$/taskMedian sToolsEvidence
Claude 4.1 Opus76.8$0.32853.13.8LIVE_MEASURED
Claude 4.1 Sonnet71.4$0.05836.03.4LIVE_MEASURED
Grok 4 Heavy71.2$0.15069.12.8LIVE_MEASURED
GPT-570.1$0.08139.83.0LIVE_MEASURED
Gemini 2.5 Pro68.8$0.02926.82.9LIVE_MEASURED
Grok 467.5$0.04831.73.2LIVE_MEASURED
o4-pro66.2$0.695114.12.2LIVE_MEASURED
Mistral Large 364.6$0.02230.92.9LIVE_MEASURED
DeepSeek V3.163.4$0.005125.02.7REPRODUCED
Qwen3-Max62.1$0.007029.62.8LIVE_MEASURED
Llama 4 405B61.2$0.007028.42.6REPRODUCED
Gemini 2.5 Flash59.4$0.001610.32.2LIVE_MEASURED
GPT-5 Mini58.6$0.004415.72.4LIVE_MEASURED
Grok 3 Mini54.1$0.002512.32.1LIVE_MEASURED
Llama 4 70B52.6$0.00165.92.0REPRODUCED