This is insane if true - could be super useful for data extraction tasks. Sounds like we could be talking in the cents per millions of tokens range.