OpenAI has begun experimenting with a pricing model that bills enterprise customers only when artificial intelligence completes tasks successfully, rather than charging for the computational resources used regardless of results. The New Stack reported that the company is already testing this outcome-based approach with select enterprise clients, marking a departure from the standard token-counting model that has dominated AI pricing since the technology's commercial debut. The shift presents both technical and financial challenges for developers building AI agents, since determining when a task truly counts as complete is far more complex than tallying tokens consumed.
Some outcomes are straightforward for software to confirm—a support ticket closing without human intervention, for instance—but many assignments leave space for interpretation. According to the report, a coding agent instructed to repair an authentication bug might rewrite the code and pass every test, only to trigger a different problem once deployed to production. In such cases, the agent technically finished the assignment, but customers probably wouldn't view it as a successful outcome. Under outcome pricing, an agent that completes 90% of a task may not generate a billable charge. The same difficulty emerges with agents that interact with browsers, databases, APIs, and other systems over extended periods—an agent might execute nine of ten steps before failing on the final one.
The report finds that evaluation tools already used to catch problems with models and agents will need to function as billing infrastructure under outcome pricing. OpenAI's hosted tools can verify responses against expected results and assess a model's performance on a given task, while platforms like Braintrust log model calls, retrievals, and tool invocations in a trace and then score the run on dimensions including task completion, factual accuracy, and correct tool use. Some results are unambiguous—a unit test either passed or didn't, an API delivered the expected response, or a database holds the record it was meant to create. Semantic evaluations are different because they demand judgment rather than pass-fail verdicts, and using that judgment to trigger a charge creates risk: a false positive could leave a customer paying for incomplete work, while a false negative could leave the vendor absorbing the cost of a successful run.
The mechanics behind outcome pricing alter who bears the cost when an agent fails. Under token pricing, the report explains, an agent can consume tokens and retry failed steps without ever finishing the job, yet the customer still pays for that usage. If customers pay only for successful outcomes, failed runs become the provider's expense—an agent that completes a task on its first attempt is more profitable than one requiring 20 model calls and several retries. The challenge grows when subjective work is involved: asking an agent to produce a monthly sales report yields something verifiable, but asking it to produce a good one means someone still has to decide whether the result meets that standard. When an AI company operates the agent and sets the success criteria, it's effectively grading its own work and then billing the customer for the result. There's also the question of responsibility when something outside the agent fails—an agent might handle a support request correctly only to encounter a timeout in the customer's CRM, or a coding agent could finish its work but fail because a separate service is unavailable, potentially leaving the vendor paying for failures it didn't cause.
OpenAI hasn't publicly disclosed pricing details or exactly how it determines when a particular task counts as a success. Still, the approach signals a willingness to absorb more operational risk in exchange for aligning billing with customer satisfaction. The model's viability will depend on whether AI systems become reliable enough to complete tasks consistently on the first attempt, reducing the provider's exposure to unpaid compute costs. For companies building AI agents, the shift means investing heavily in evaluation infrastructure that can distinguish genuine success from near-misses, since the line between a billable outcome and a costly failure will determine profitability. This pricing experiment may ultimately reshape how organizations budget for AI—not as a metered utility but as a service purchased by the job, with all the accountability that comes with delivering results rather than simply running models.