Key Takeaways

  • OpenAI considered dropping token-based billing entirely for a flat quote per prompt, but enterprise customers pushed back because they demanded an auditable, quantifiable unit.
  • Token counts in modern AI models have become an abstraction similar to semiconductor nanometer marketing, yet buyers still prefer them over black-box pricing.
  • Stripe currently spends about 1% of its total operating expenses on tokens, and John Collison expects that figure to climb to 5%, 10%, and eventually 20%.
  • Sam Altman argues that model selection and reasoning step overhead waste founder time, creating pointless friction between 10-cent small models and high-end options.

The Customer Revolt Against Black-Box Pricing

OpenAI almost killed the token.

When planning their model rollouts, Altman considered junking token-based pricing in favor of dynamic, prompt-level quotes. You would submit an input, the API would return a price, and you would decide whether to run it. It sounded cleaner on paper. In practice, developers hated it.

“We almost made a very, what would've been bad call to say we're not going to price our next model in tokens at all,” Altman explained. “And you're just going to give it a prompt. And we're going to give you a price back.”

Developers wanted an underlying physical anchor, even if that anchor was mostly an illusion. Altman pointed out that token counts today mirror semiconductor nanometer measurements. When chip foundries market a "3nm" process, the gate length does not actually measure three nanometers anymore. It is a marketing abstraction. Yet software teams still demand that abstraction over pure vendor discretion.

Altman noted that customers made their stance obvious: “Even if we inflated the tokens and they're like, 'Oh, I'd much rather you inflate the tokens and pay more. Because I know there's a real thing there.'”

Engineers will accept arbitrary markups before they accept unmetered bills. When you run infrastructure at scale, an imperfect metric you can audit beats a opaque price tag you cannot verify.

Stripe Wants to Spend 20% of OpEx on Models

While early-stage teams obsess over lowering their API bills, the largest tech companies are trying to spend more.

Collison revealed that Stripe allocates approximately 1% of its operating expenses to tokens. In his view, that number represents an underinvestment in automation. Stripe is deliberately pushing that spend upward.

"That feels too low," Collison said. “So, we're working on fixing it. And it'll get to 5%, and 10%, and 20%.”

If a company with thousands of employees shifts 20% of its operating budget from payroll and software licenses to tokens, the nature of its labor force changes. The spend replaces headcount growth, routine support contracts, and manual fraud review. Treating model inference as a cost center to trim is a mistake when the output directly replaces more expensive operational work.

Altman echoed this perspective. Founders often spend hours debating whether to route tasks to a cheap model at ten cents per million tokens or spend ten dollars per million tokens on a heavyweight reasoning engine. That optimization costs more in human engineering time than it saves on the monthly cloud invoice.

“Right now you have to think about whether you should want the 10-cent-per-million tokens out of a small model or $10 per token tokens out of a big model,” Altman said. “And then whether, how long the chain of thought is likely to be. And that's just cognitive overhead that I don't think actually adds any value.”

What to Do With This

Audit your team's model routing logic tomorrow morning. If senior engineers are spending days benchmarking small open models to save $400 a month on support classification, kill the project and route everything to the best frontier model available. Calculate your token spend as a percentage of your total operating expenses; if it sits below 2%, you are under-automating routine operations that your competitors will soon turn over to software.