OpenAI is testing pricing where some customers pay only for finished tasks

The Information reports that OpenAI has started giving some major customers the option to pay only when its AI actually completes a task, such as handling a customer support conversation. The outlet published the briefing on Sunday.

Two things to be clear about before going further. This is that outlet's reporting, and their story is behind a paywall, so the detail below is limited to what they stated. OpenAI has not published anything about it, and we have not seen a contract or a price list. Treat it as a credible report rather than a confirmed product change.

The Information also frames it as part of a wider shift, noting that Salesforce and other companies have been moving toward charging for outcomes rather than usage.

Why this is a bigger change than it looks

Almost every AI product today bills for input. You pay per token, per call, or per seat. The customer pays the same whether the model got it right or produced something unusable.

Paying for completed work moves that risk to the seller. If the model fails a support conversation four times before succeeding once, the four failures reduce the seller's profit. That is a real change in who pays the cost of a model being wrong.

Who gets to define completed

Two things follow, and they matter whether or not you ever buy from OpenAI.

The first is that outcome pricing only works if both sides agree on what "completed" means. Handling a support conversation is not a yes or no fact. Somebody has to define the check, and that definition becomes the commercial contract. A vague definition will likely lead to a dispute.

The second is that this makes evaluation a commercial requirement instead of an engineering preference. If revenue depends on a task being finished correctly, you need a reliable way to decide whether it was. That is an eval suite, whatever it gets called in the contract. A team selling on outcomes without one is guessing at its own revenue.

There is a version of this that goes wrong, and it is worth stating. When the seller writes the success check, the seller has an incentive to write a check that is easy to pass. Any buyer signing an outcome-based deal should ask who defines success, who can audit it, and what happens when the two sides disagree.

For teams selling AI products to other companies, this is a sign of what buyers will expect. Charging per token asks a customer to pay for your model's mistakes. Once a large seller stops doing that, it becomes harder for smaller sellers to keep doing it.

Source

• The Information, "OpenAI Starts Letting Some Customers Pay Only When the AI Works"