There’s been some recent buzz in the news about OpenAI migrating to “outcome-based pricing”, with many outlets calling this out as a transition from the standard token-based pricing model available via the API. Signs point to a different development. We’ll share why and explore how outcome-based pricing works and can be scaled effectively.
If you’re familiar with AI customer-support automation, then you’re most likely familiar with outcome-based pricing already. Providers, including Valiopt, charge a fixed fee per resolved or automated support interaction (though the definition of said interaction may vary). Basically, if the AI agent can resolve the customer support request, you pay; if it’s escalated to a human agent, you don’t.
This type of outcome-based pricing aligns incentives and can be a great way for software providers to demonstrate real value compared with simply selling a fixed-price subscription. It’s in the service provider’s best interest to resolve more conversations, which is typically what the customer is trying to do, since the per-outcome pricing is cheaper than having a human handle the task.
Outcome-based pricing works well in customer support because the outcomes are well-defined, and the variation in cost and complexity in resolving different types of requests is typically low. Customer support requests change little over time and require only a few turns to resolve.
OpenAI recently launched OpenAI Presence, a managed enterprise service for this type of work. While few details about Presence are publicly available, it seems to be the same shape of service as what’s provided by a number of customer support AI players (Sierra, Intercom Fin, Valiopt, and others). OpenAI configures AI agents that answer questions, interface with third-party systems, and are designed to handle inquiries end-to-end.
Our suspicion is that recent reporting about OpenAI switching to outcome-based pricing is referring to enterprise service arrangements such as OpenAI Presence, as opposed to a real transition from token-based pricing.
Moving the OpenAI API from token-based to outcome-based pricing would be difficult because requests vary widely in their complexity and in how easily an outcome can be measured. Asking ChatGPT to generate an HTML website could use 1 million tokens, or it could use 1 billion. At the end of the interaction, it’s also not always clear whether the user achieved the desired result. Measuring success isn’t as easy as determining whether a support request was escalated to a human.
Outcome-based pricing is a great way to align incentives between software providers and their customers, and it works incredibly well for tightly defined tasks and processes. The recent coverage does not appear to signal a shift in how frontier model providers like OpenAI are thinking about pricing; rather, it suggests that they’re expanding their product offerings to include more customized solutions that mimic what third parties offer in the customer support space.