bias against invoking LLMs is maybe a result of token/task benchmaxxing?