More confusingly, they both use different tokenizers so token counts aren’t 1:1 at all.