Hacker News Reader

He's saying output for 1M tokens on the latest models is $50 now when it used to be $2500.

so how are these labs going to recoup the insane training costs at those prices? even if there is still a fat margin leftover afterwards

They also have to continuously train, forever, to avoid model drift. It's not a one and done thing as far as I'm aware.

Diminishing returns without some major efficiency gains.