You're touching on what I think is the hardest unsolved problem in this transition. At DataSift we went through something similar - pricing firehose data access when the value was in the insights, not the volume. We ended up with a hybrid model.

My bet is we'll see "outcome pricing" win for most use cases - you pay for "site deployed" not "tokens consumed." But that requires the provider to absorb variance in AI costs, which is scary when models are still getting cheaper quarterly. The providers who figure out how to price on outcomes while hedging their compute costs will win big.

What pricing model is working best for you so far?