OpenAI came close to abandoning token-based pricing for its next model in favor of a flat-fee structure where developers would submit a prompt and receive a single price, before customer pushback forced the company to reverse course, CEO Sam Altman revealed. Speaking Thursday on the Cheeky Pint podcast with Stripe President John Collison, Altman described the proposal as "what would've been a bad call" and said the billing would have worked like Amazon Web Services, where "you just do something and we send you a bill" instead of measuring each exchange by tokens. Altman didn't identify which model was involved or when the change was under review.
The plan collapsed when OpenAI surveyed customers about getting a price without any itemized breakdown. Altman said the reaction was "so negative," with buyers insisting they needed to understand what they were paying for. Developers already invest significant effort monitoring input and output tokens to control expenses and project future costs as their usage scales, and receiving an unexplained charge would only complicate that work. What made the disclosure particularly striking, according to Altman, was that the people he consulted said they'd accept higher rates to preserve visibility into how charges were calculated. When he asked "a few people" whether they'd pay a premium to keep consistent token-based pricing for the same output, they told him they'd "much rather you inflate the tokens and pay more" because "I know there's a real thing there."
Altman explained that developers resist a single opaque price because they fear they'll "get screwed somehow" over time, the same wariness that once led buyers to purchase CPU hours rather than finished jobs. The proposed shift could have carried major consequences for developers building agents, whose multi-step workflows already burn through unpredictable volumes of compute and make spending hard to forecast even under current token metering. Altman didn't clarify how OpenAI would have determined per-job prices or whether customers could have compared one charge against another.
The topic surfaced when Collison challenged Altman on whether tokens were the appropriate metric for measuring AI usage. Altman conceded that "tokens are a kind of fake unit," adding "they're real in some sense, but they're not a good unit," though he offered no concrete alternative. He rejected Collison's half-joking suggestion to create an intelligence-equivalent token, likening it to how chipmakers still advertise nanometer specifications that no longer match any physical dimension on the chip. Altman acknowledged that developers face genuine friction when weighing a cheaper model against a more powerful one, especially when they must also estimate how much reasoning a task will require, calling it "cognitive overhead that I don't think actually adds any value." The conversation arrives as OpenAI continues using token pricing even as Altman has contended that the cost of any given AI capability drops roughly tenfold annually, and for now the company keeps tokens in place despite Altman's admission of their limitations and OpenAI's struggle to discover a better option. The episode lays bare a tension every AI platform will eventually confront: simplified billing may suit the vendor, but opacity around costs can erode the developer trust that underwrites long-term adoption.