OpenAI tests pay-per-success agent pricing

This is the pricing model agents eventually need, but “successful completion” becomes difficult to define once workflows span multiple tools, retries, and human handoffs.

• –Vendors absorb the cost of failed attempts, creating stronger incentives to improve reliability and evaluation.
• –Enterprise buyers gain more predictable ROI, while OpenAI must price in inference, tooling, retries, and support costs.
• –Customer support is an easy starting point because resolutions can be independently verified.
• –Developers building agents will need explicit success criteria, observability, and audit trails before outcome billing can scale.
• –The selective rollout matters: OpenAI’s broader enterprise plans still support credit- and token-based pricing, so this is an experiment rather than a general replacement.