Michael JohnsonWhen an application needs to choose one of several next steps, a full generative response can add unnecessary latency and expense. On October 22, 2026, OpenAI announced the general availability of Decisions API, an interface for constrained decisions where developers specify the question and allowed answers in advance.
According to the company, the API is built on GPT‑6 Luna and accepts text, images, and structured data. OpenAI says it works up to 10 times faster for its target tasks than GPT‑6 Luna through the Responses API. That comparison applies to the specified processing route; you should test performance in your own system using real inputs, including classification errors and the share of requests that require manual review. Details are available in the Decisions API announcement and the developer guide.
Three output types instead of free text
Developers describe the decision and its context, and the API returns one of three output types. Predicate estimates the probability of a condition, such as whether visible damage is present. Choice selects one option from a predefined list—for example, continue, retry, or request clarification. Score evaluates the input against an ordered scale defined by the developer.
This format can be used to classify support requests, route documents, choose an agent’s next action, and perform an initial risk assessment. Its key practical feature is that the application defines the limits of the response: the result can be connected directly to code branching, while low-confidence cases can be routed to a human. In its announcement, OpenAI cites Box as an example: the company said that, in its own evaluation of 30 insurance certificates, Decisions API selected a route about five times faster than Responses API at the same accuracy. That result comes from a specific Box test and is not a universal guarantee for other datasets.
What to check before deployment
OpenAI lists pricing at $0.10 per million input tokens and $0.01 per million cached input tokens; there is no charge for output tokens. Up to 400 questions are supported in a single request. The company also lists caching for reusable context, data processing in the US and EU, and Zero Data Retention availability for eligible customers.
To assess the economic impact, choose one frequently recurring decision point and compare the API with your current implementation on saved examples. Record accuracy for each class, latency at the 50th and 95th percentiles, the share of uncertain cases, and the cost of the full request, including context. For consequential tasks—such as automatically executing a command or declining an operation—add a confidence threshold and a human-review path for borderline responses.
Also check the quality of the instructions and the list of allowed options: if an important case is missing from the answers, the model has to choose an unsuitable outcome. For voice applications, OpenAI announced built-in integration of Decisions API with GPT‑Live; the company gives the example of choosing the next action during a live conversation. For this scenario, it is useful to measure latency from the arrival of the user’s signal to the interface’s response.
Where the new API could be useful
Decisions API targets small decisions that recur many times within a larger process. That makes it a candidate for initial request triage, routing tasks between queues, evaluating incoming documents, and choosing an agent’s next step. The practical adoption test is straightforward: the API should improve latency or cost on your task while maintaining acceptable accuracy and providing a clear way to handle uncertain cases.
For teams already using Responses API, a sensible first step is to extract one constrained decision point rather than rewrite the entire process, then compare it on identical examples. This lets you test whether a fast structured response really simplifies the system and justifies a separate processing path.