OpenAI has started billing a small set of its largest customers only when its AI finishes the task, not when it merely attempts it. The Information’s Kevin McLaughlin and Amir Efrati reported the shift, citing a customer support scenario handled start to finish. OpenAI has not announced the program, and the customers, terms, and prices involved remain unknown.

That caveat matters more than usual here. This is a single report describing a selective, unannounced arrangement, not a confirmed policy change, so treat every detail below as provisional until OpenAI says otherwise.

The pressure behind it is not speculative, though. Token pricing has turned into an accounting headache once agents run unattended: a single developer with a hundred parallel agents burned $1.3 million of OpenAI tokens inside a month, a number that advertised the problem rather than creating it. When cost tracks attempts instead of results, a finance team cannot forecast the bill, and an unforecastable bill is the easiest reason to delay signing.

Outcome pricing already has a working model to borrow from, in customer support specifically, largely because a resolved ticket is a rare case where everyone agrees what finished looks like. Intercom bills only on the conversations Fin actually closes, and waives the rest. Zendesk narrowed further in May, billing only for what it calls Verified Resolutions, confirmed by a separate LLM evaluation within 72 hours, at roughly $1.20 to $1.50 per resolution on committed volume.

Salesforce shows what happens when a vendor gets the definition wrong before fixing it. Agentforce launched charging $2 for every 24-hour conversation session regardless of outcome, which customers found expensive and impossible to plan around. Salesforce’s answer, Flex Credits, bills by individual action at about 10 cents each rather than by outcome, which is consumption pricing wearing outcome pricing’s reputation.

Buyer preference research backs the shift in direction, if not yet in magnitude: Futurum Group found in May that 43% of enterprise buyers prefer consumption-based billing and 27% prefer outcome-based billing, versus under one in five who still want to pay per seat. Keith Kirkpatrick, who directs enterprise software research there, wrote that a seat-only price list now gets a vendor cut before the evaluation even begins.

None of these support-desk examples resolve the harder case OpenAI is reportedly stepping into. A support ticket either closes or it doesn’t. A multi-step agentic task, the kind OpenAI has been pushing toward with the 10 million users it says are on its agent products, has no equivalent binary marker. Someone still has to decide, task by task, whether the output counts as done.

That decision is the actual product being sold here. The hard part of outcome pricing was never the billing mechanics; it’s writing the definition of “worked” precisely enough to hold up in a dispute, and whoever controls that definition, the vendor grading its own homework, sets the effective price and captures the margin regardless of what the headline rate says.

For any team currently negotiating an OpenAI enterprise contract, the question worth raising now is who defines completion and what happens when that definition is contested, before it becomes a term buried in a renewal.

Reporting via The Next Web (TNW), based on a story first reported by The Information, published August 31, 2026.