I think these volumes are relatively small. Bear in mind that this is tokens, not API calls. Depending on how much input text or output text there is a single call could easily be using 100s of tokens.