Priced by the Token
Not by characters, not by time, not by number of calls, but by a unit called the token. It began with the OpenAI API, opened to invitations in June 2020, and every AI company now counts the same way. Prices kept falling: Davinci went from six cents per thousand tokens to two in 2022, and gpt-3.5-turbo arrived the next year at $0.002. An order of magnitude at a time. Two things make this new in the long history of metered pricing. First, the unit cannot be counted by the person paying — unlike Xerox's page, a token is an internal slice the machine makes of a sentence. Second, you are billed for what you cannot see: the draft a reasoning model writes to itself is charged at the same rate as its answer. Paying for the process rather than the result — the river of «only for what you used», running since the copier of 1959, has begun for the first time to count something invisible. Consumers, incidentally, pay a flat twenty dollars a month (from February 2023). The same product metered to companies and fixed for individuals — a two-storey arrangement this map has met many times before.
The myth, corrected
AI APIs were priced per thousand tokens from the very start
They began as a monthly plan with overage. At launch in 2020 it was $100 a month with two million tokens included and eight cents per thousand beyond that; the clean metered pricing we know came later. Even the newest trade in the world did not settle its pricing on the first try.