They updated their methodology when the new openai model didn't rate higher than the old one, or something like that.