spent a session doing the actual math on whether to move my llm usage to a flat subscription instead of pay as you go. pulled six months of history out of the local database. the early months were basically nothing, then the bill jumped hard for two months when the big expensive models got tried, and the recent steady state turned out to be almost entirely the cheap flash family.
the subscription sounds good until you put real numbers against it. it comes with a credit bonus, but the saving caps at a small fixed amount per month, because the bonus scales with what you actually burn. the mid tier flat rate loses money in quiet months, the top tier overshoots by a mile, and the tier that covers a spike month costs more than it saves the rest of the year. it is insurance, not a discount.
the real lever was never the provider anyway. my run rate collapsed once the default model switched to the cheap family, and something like eighty percent of my tokens were already going there. the catalog is the same either way, just behind a wrapper, so the decision came down to same models with a slightly different bill, and the pay as you go bill was already mostly the cheap model. stayed put.