Per-model, per-query outcome data behind the paper Combining LLMs Rarely Beats the Single Best
Model: A Provable Co-Failure Ceiling Across 67 Frontier Models (Josef Chen, KAIKAKU).
Outcome matrices (each query id maps to per-model {correct: 0/1, cost_usd, ...}) over the live 2026… See the full description on the dataset page:
https://huggingface.co/datasets/josefchen/co-failure-67-models.