A retrieved chunk arrives from the search rather than from you, so it feels like something the model produced. It is not. Retrieval puts the document into the request, which makes every chunk an input token: priced as input, and re-read on every turn after the one that fetched it. So a document retrieved once is paid for again and again, long after the question that went looking for it.
Paid members: the worksheet and its printable PDF are below ↓