Results

Each run is a fixed historical study with its own protocol, evidence and limitations. A model does not possess one timeless CookingBench rank.

Archived preliminary

2026-07-v2.1

29 July 2026

CookingBench v2.1

14 models, 184 prompts each and 2,576 preserved responses. The original published scores remain visible, but the scoring instrument did not support a unique fine-grained ordering of the leading models.

View archived result Read the autopsy Open corpus record