For years, buying software meant paying a fixed price whether you used it a lot or barely touched it. That model is breaking down for AI, and OpenAI just joined the shift.
According to The Information, OpenAI has begun offering some large customers a new deal: pay nothing until the AI actually finishes the job, such as resolving a customer support request on its own. OpenAI has not confirmed this publicly, and the change was not announced.
This matters because OpenAI is the biggest name in the industry, and its move signals that pay-for-results pricing is becoming the norm, not a niche experiment. Sierra, a customer service AI company, already charges this way. So does Fin, which Salesforce is buying for three point six billion dollars.
Cognition, which makes the coding assistant Devin, goes even further. It will hand back up to ten million dollars in credit to enterprise customers if the tool fails to deliver value worth at least what they paid.
The money problem behind this is real. Running AI models is costly, and that cost has been climbing fast.
OpenAI's own cost of serving its products nearly tripled in a single year. It grew from about two point six billion dollars in 2024. By 2025 it reached roughly seven point five billion dollars.
There is real precedent for this outside of software. Rolls-Royce started charging airlines by the hour of jet engine use back in 1962, rather than selling the engine outright, a model still known as Power by the Hour. AI vendors are simply applying the same logic to office work.
The catch is proving whose work actually produced the result. If a support ticket closes or a sale happens, was it the AI, a marketing campaign, or just a good month? Stripe, which processes payments for many of these vendors, has already published guidance warning that pricing on outcomes you cannot independently verify leads straight to billing disputes.
Analysts at Gartner expect a fast rise in this kind of pricing. By the end of 2026, roughly four out of every ten enterprise software deals are expected to include some form of outcome-based pricing. Two years earlier, that share was only about fifteen percent.
For any business buying AI tools, this is good news in theory: you stop paying for AI that does not work. But it only helps if the contract spells out exactly what counts as success before you sign. Otherwise you are trading a predictable bill for a fight over credit every time something goes wrong.