{
  "id": 4698198,
  "title": "OpenAI wants to charge only when AI gets it right — here’s the catch",
  "url": "https://urgent.news/2026/08/31/openai-wants-to-charge-only-when-ai-gets-it-right-heres-the-catch",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-31T18:47:25.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/openai-outcome-based-pricing/"
  },
  "original_language": "en",
  "account": "Traditionally, AI companies bill customers based on the tokens consumed by their models, regardless of whether the AI delivers the desired results. OpenAI is exploring a novel pricing model, permitting customers to pay exclusively when the AI successfully completes a task. This approach, first reported by The Information, is currently being trialed with select enterprise customers. The specifics of how OpenAI determines when a task counts as successful remain undisclosed. This innovative billing method poses a technical challenge for developers constructing AI agents. Establishing when a task is deemed complete is more intricate than merely counting tokens. While some outcomes are straightforward to verify, such as closing a support ticket without human intervention, other tasks are open to interpretation. For instance, an AI agent tasked with fixing an authentication bug might generate the corrected code and pass all tests, yet the patch could introduce new issues when deployed in the production environment. In such cases, the agent may have technically completed the task, yet the customer would likely not deem it a successful outcome. Outcome pricing introduces a new complication, as the completion of 90% of a task may not suffice for the run to be considered billable, unlike with token pricing where 100% activity usage is charged. Developers utilize evals to identify issues with models and agents. OpenAI offers hosted tools to assess responses against expected results and grade a model's performance on a given task. Braintrust provides visibility into the intricacies of an agent run, recording model calls, retrievals, and tool calls in a trace and scoring the run based on factors like task completion, factual accuracy, and correct tool usage. These traces can be transformed into datasets for future testing. Certain evaluations, like unit tests, are straightforward—either they pass or they fail, exhibiting minimal room for debate. Others, such as semantic evals, necessitate human judgment instead of a binary pass/fail. An LLM-as-a-judge aids developers in comparing two iterations of an agent, yet translating that judgment into a billing mechanism presents a distinct challenge. Incorrect judgments could result in customers paying for incomplete work or providers covering the costs of successful runs. When AI companies conduct the agent and establish the criteria for success, they are effectively grading their own work and subsequently billing the customer for the outcome. This process becomes increasingly complex when dealing with subjective tasks. Generating a monthly sales report is verifiable, but instructing an agent to create a high-quality report raises the question of how to ascertain the result's quality. Evaluating the improvement in conversion rates attributable to the agent is even more challenging. Moreover, attributing responsibility when external factors impede an agent's performance adds another layer of complexity. An agent may execute a support request accurately but encounter a timeout in the customer’s CRM. A coding agent might successfully complete its task but fail due to an unavailable separate service. In such scenarios, the vendor could incur costs for failures beyond its control. Failed agents consequently impact profit margins. Under token pricing, an agent can waste tokens and retry failed steps without ever concluding the job, with the customer still paying for the usage. In contrast, outcome pricing shifts responsibility for failed runs to the provider, as they would have spent money on compute resources without generating any billable output. Agents that succeed on their first attempt generate greater profits than those requiring numerous model calls and retries. If an agent never concludes, the provider bears the financial burden of the failed computation.",
  "summary": "AI companies have always charged customers for the tokens they use, whether the model gives them exactly what they need The post OpenAI wants to charge only when AI gets it right — here’s the catch appeared first on The New Stack .",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}