4 ms·

A large part of the GPT-5.x model iteration has been about making training more affordable and token efficient.