OpenAI wants to charge only when AI gets it right — here’s the catch

OpenAI is trying a very different billing model with some enterprise customers: pay only when the AI actually finishes the job. The idea was first reported by The Information, and it swaps the old token meter for something closer to a results meter. If the system succeeds, OpenAI gets paid. If it doesn’t, the company eats the cost.

That sounds clean until someone has to define success. OpenAI hasn’t said how it sets the price or how it decides that a task is done well enough to bill for it. And that is where the neat idea starts rubbing against reality, because AI agents often do most of a task and still miss the part that matters.

Some work is easy to score. A support ticket is closed. A unit test passes. A database record shows up where it should. But a coding agent that patches an authentication bug can still ship a new problem later in production. The code might look finished on paper and still fail the person paying the bill.

The mess gets worse with agents that move through browsers, databases, APIs, and other systems. They may complete nine steps and die on the tenth. Under token pricing, all that failed effort still gets billed. Under outcome pricing, the provider may have spent compute on a run that never becomes revenue.

That is why evals are turning into billing infrastructure. Tools from OpenAI and others can already grade task completion, factual accuracy, and correct tool use, and Braintrust logs model calls, retrievals, and tool calls in a trace. But a judge that helps developers compare agents is not the same thing as a judge that decides who gets charged. False positives mean customers pay for unfinished work. False negatives mean the vendor pays for a success.

The real trap is subjective work. A monthly sales report can be checked. A good monthly sales report is another story. And once the result depends on outside systems, like a CRM timeout or a separate service going down, the blame can land in the wrong place fast.