About extractingbench

extractingbench is an independent benchmark of extraction models, run by Leo. It sits next to scrapingbench (web data APIs for agents).

Rules

Every number on the site is read from the results file by code. None is typed by hand.
No model is called the winner unless the 95% intervals support it.
No numbers are published for a model we haven't tested. It is listed as not tested.
Each provider's terms are read before its results are published. A provider that forbids publication is skipped and named.
No vendor or sponsor has any say in the method or the results. No ads, no affiliate links, no tracking scripts and no cookies.