Or switch to using the way cheaper open weight models from various providers who don’t have to subsidize training costs so can just race to the bottom on inference pricing…

The quality isn’t really SOTA yet but at some point I assume they’ll be good enough (maybe already are?).