I’m pretty sure input tokens are cheap because they want to ingest the data for training later no? They want huge contexts to slice up.

Afaik all the large providers flipped the default to contractually NOT train on your data. So no, training data context size is not a factor.