OpenAI has started letting some customers pay only when the AI works

OpenAI has begun letting some of its largest customers pay only when its AI actually completes the job. Kevin McLaughlin and Amir Efrati reported the change to The Information, giving the example of a customer support interaction handled end-to-end.

The arrangement is limited to select major accounts rather than offered generally, and OpenAI has not announced it. TNW has not independently verified the report, and the terms, the customers, and the prices are all unknown.

The industry name for this is outcome-based pricing, and the appeal to a finance director is not subtle. A bill that arrives only when something worked is considerably easier to defend than one that arrives regardless.

Token billing has made that a live problem. One developer running a hundred agents in parallel accumulated $1.3m in OpenAI tokens across thirty days, which is an extreme case of a general pattern where cost scales with attempts rather than results.

Customer support is where the model has already settled, because a resolution is one of the few AI outputs anyone can define. Intercom charges $0.99 for each conversation its Fin agent resolves, and nothing at all for the ones it does not.

Zendesk went further in May, restricting billing to what it calls Verified Resolutions, confirmed by an LLM evaluation within 72 hours of the conversation. Assisted escalations and contained resolutions became free, and the billable rate sits at roughly $1.20 to $1.50 on committed volume.

Salesforce has been working through the same question in public and more awkwardly. Agentforce launched at $2 per conversation, charged for every 24-hour session whether or not anything was resolved, which customers found both expensive and impossible to forecast.

Flex Credits arrived as the answer, moving the meter from conversations to individual actions at about 10 cents each, starting at $500 for 100,000 credits. That is consumption pricing rather than outcome pricing, and the distinction matters, because an action that fails still bills.

Buyers appear to want both, in that order. Futurum Group found in May that 43% of them prefer consumption-based models and 27% prefer outcome-based ones, with fewer than one in five still preferring to pay per user.

“Outcome-based pricing is becoming a market standard,” wrote Keith Kirkpatrick, the firm’s research director for enterprise software, whose sharper finding is that vendors offering seats alone are now being disqualified before the evaluation starts.

For OpenAI, the move is a change of position rather than a new product. It has sold capacity by the token, priced per model and per call, and letting an enterprise pay for completed work instead means accepting the risk that the work does not complete.

That risk has to be priced somewhere, and the interesting question is where. A vendor confident in its success rate can afford the arrangement, while one that is not has to load the per-success price until the economics match, which is why per-resolution rates cluster around a dollar rather than a cent.

It also changes who carries the cost of a bad answer. Under token billing, the customer pays for every failed attempt, whereas under outcome billing the vendor absorbs them, which is a meaningful transfer of risk from the buyer to the company that built the model.

The harder part is agreeing what success means. A resolution is definable, but the agentic work OpenAI has been pushing towards, with 10 million users on its agents, involves multi-step tasks where completion is a matter of judgement rather than a field in a database.

There is a commercial reason to want it settled quickly. An enterprise that cannot forecast a bill tends to run a pilot indefinitely rather than sign, and outcome pricing removes the objection at exactly the point in the sales cycle where it usually stalls.

None of that is settled, and none of it is public. What is on the record is that the largest model vendor has started, quietly and selectively, to sell results instead of capacity.