OpenAI has taken its Decisions API out of public beta, unveiled during the company’s DevDay on September 29. Designed for applications and agents that chain together many small decisions, it does not write text but returns a typed response, fitting into the niche of so-called decision models.
An API that decides instead of writing
OpenAI starts from an observation it summarizes as follows: “Applications and agents make many small decisions throughout a workflow. Individually, they are simple. At scale, their latency and cost add up.” The operation of a decision model remains simple: the developer formulates a question, defines the allowed answers, and provides the context, in text or image form. The API then returns a response in one of three formats, which the application uses to trigger the next step.
The Predicate format gives the probability, between 0 and 1, that a condition is true, for example the presence of visible damage on a product photo. The Choice format selects an option from a list set by the developer, with probabilities and a confidence score. Finally, the Score format assigns a rating on a grid of ordered levels, such as the severity of damage.
A single query can now group up to 400 questions, twice as many as in beta. They all relate to the same content and are evaluated independently, so a question depending on another answer requires a separate query. Technically, OpenAI has not trained a specific model: the API relies on GPT-6 Luna, reconfigured to accelerate inference, i.e., the processing of a query by the model. According to the company, it focuses “Luna’s intelligence on questions with bounded answers” and makes its decisions up to 10 times faster than Luna when queried via the Responses API.
JSON, cache, and voice on the menu for the final version
Since the beta, the API accepts inputs in native JSON, a structured data format, as well as images embedded in a question’s instructions. Prompt caching reuses the common part of a query, a reference document for example. Inference processing can now be done in the United States or the European Union, in addition to the Zero Data Retention offered to eligible customers. The API also integrates natively with GPT-Live, OpenAI’s model dedicated to voice conversations: a customer support agent can thus decide, mid-conversation, to ask a clarifying question or transfer the request to a human.
Regarding pricing, the Decisions API is billed $0.10 per million input tokens, and $0.01 per million cache tokens. Output tokens are not billed. Writing to the cache, for a duration of 30 minutes or 12 hours, is subject to separate pricing, and regional processing may incur an additional cost.
A principle close to Jev
The operation is reminiscent of Jev, the decision model from startup TypeSafe AI. The same three question types are found, OpenAI’s “predicate” corresponding to Jev’s “Noul,” as well as billing limited to input tokens. The approach differs, however: TypeSafe AI designed a dedicated architecture and training method, whereas OpenAI adapts an existing language model. The Decisions API also costs more than twice as much as Jev, which was billed $0.042 per million input tokens at launch. Support for images, however, no longer distinguishes the two offerings, as TypeSafe AI added it to Jev in its latest version.
According to two independent rankings dedicated to decision models, the AIM System One Index and the Artificial Analysis Decision Index, OpenAI’s model would rank behind Jev, with both dominated by Liquid’s D1 Ultra and Base Lab’s Elephant-1. OpenAI offers to test its API on examples from real-world uses in its Decisions Playground and provides the necessary documentation for developers.
