It still happens even when saying explicitly these are real numbers for ChatGPT 4o (see below). Also it seems to critically depend on the word order — I've only seen the model fail when the choices precede the question.

→ View original post on X — @goodside