OpenAI’s Decisions API handles a narrower task than an AI agent. Given a question and a finite set of developer-defined answers, it selects one using text or image context. OpenAI is pitching it for classification, request routing, and choosing an agent’s next action.
A support system, for example, could ask which team should receive a request instead of asking a model to compose an unrestricted plan. The application still has to decide what to do with the selection. Limiting the answers does not guarantee it will be right.
The API is in limited preview. At its September 29, 2026, announcement, OpenAI said it planned a broader release in the coming days. Its initial materials leave questions about cost, performance, and access unanswered.
Developers Define the Decision Before Luna Makes It
In its DevDay 2026 recap, OpenAI describes Decisions API as using Luna to choose among developer-defined answers based on text or image context. The developer supplies the question and possible answers; Luna selects from that set. OpenAI’s developer account calls the underlying model GPT-6 Luna.
Consider a hypothetical support workflow. A developer might ask, “Which team should receive this request?” and offer “billing,” “technical support,” “account access,” and “manual review” as possible answers. The request supplies the context, and the application could use the selection to send the case to a queue. These labels illustrate a bounded decision; they are not a published Decisions API schema or a demonstrated result.
According to OpenAI, a developer can also provide an image as context for a classification question. The announcement does not establish how well Luna interprets screenshots, documents, or other image types in a particular workflow. Teams would need to test their own inputs before relying on image-based selections.
OpenAI’s developer account summarizes the intended uses in its preview announcement:
Defining the available answers is different from knowing which one is correct. A restricted set can make a decision easier for an application to handle, but it does not show how the system deals with ambiguous inputs, missing information, or cases that fit none of the labels.
A Routing Decision Is Not an Agent Workflow
OpenAI announced several developer tools at DevDay. The developer community announcement describes the Agents API as a public beta with hosted execution, memory, tools, and multi-agent support. Decisions API has a different job: selecting an answer from a predefined set.
An agent might use tools to carry out a task over several steps. A decision can determine which task should start, which queue receives an item, or which permitted action an application should consider next. OpenAI lists selecting an agent’s next action among the new API’s uses, but does not describe Decisions API as executing that action.
An application could therefore seek a bounded routing choice before handing work to an agent. That is a possible design pattern, not an announced integration or a claim that Decisions API will prevent an agent from taking an unwanted action.
Conventional rules remain another option. If a request contains a reliable, structured field identifying its destination, a fixed rule may be sufficient. A model becomes more interesting when the relevant context is messy, such as an unstructured message or an image that needs interpretation. OpenAI has not published evidence showing when Luna outperforms simpler routing methods, so developers cannot yet calculate the trade-off from the announcement alone.
Fixed Choices Still Need an Escape Route
A finite answer set shifts some risk from free-form generation to the design of the choices. If every option names a department but the input is unreadable or unrelated, selecting a department may be the wrong operational outcome. Developers can include “manual review” in a proposed workflow, but that label’s presence would not prove Luna will choose it whenever a human should intervene.
The answer list is an engineering decision, not merely a prompt-writing detail. Labels need to be distinct enough to act on. Teams also need a policy for borderline cases and a way to check selections afterward.
A support-routing pilot, for example, could compare selections with cases already assigned by human reviewers. Useful measures would depend on the workflow: how often requests go to the wrong team, which categories get confused, and whether uncertain cases reach review. A single overall accuracy figure could conceal a costly error pattern if one category rarely appears but requires urgent handling.
OpenAI’s initial announcement provides none of those results. It also does not specify a confidence measure, built-in abstention behavior, or reliability guarantees for Decisions API. Developers should not assume those features exist because they would be useful for bounded classification.
Text and images can contain instructions written by someone other than the application developer. A submitted request might, for instance, demand a particular destination. Testing should include such cases to determine whether the system bases its selection on the intended evidence. OpenAI’s description establishes the API’s inputs and purpose, not a security guarantee against manipulation.
Limited Preview Leaves Deployment Questions Open
Decisions API is available in limited preview. In a follow-up to its developer announcement, OpenAI said access was limited to selected API customers for testing. Its DevDay recap said a broad release was planned in the coming days. That was a plan stated at launch, not a published general-availability date.
The supplied announcement materials provide no pricing, access criteria, or public performance results for developers outside the selected group. They also lack enough detail to judge response times, how outputs are represented, or how uncertain selections are handled. OpenAI describes a real-time decision-making use case, but gives no measured latency figure.
A team with preview access could start with one decision that has clear, finite outcomes, then assemble representative examples, including cases that should go to manual review. Comparing Luna’s selections with existing rules and human decisions would reveal more about its value than a polished demonstration. A team without access can prepare that evaluation, but should wait for documentation and pricing before planning a production migration.
Frequently Asked Questions
4 questions
1What Is OpenAI Decisions API?
OpenAI Decisions API is a limited-preview API that uses Luna to select an answer from a finite set defined by a developer. OpenAI describes classification, request routing, and choosing an agent’s next action as intended uses. The developer supplies the question and possible answers, while text or image context informs the selection.
2
Sources
- DevDay 2026 recapopenai.com
- https://x.com/OpenAIDevs/status/2105003318917697873x.com
- developer community announcementcommunity.openai.com
