Introducing the Decisions API

A fast lane for bounded choices

OpenAI’s Decisions API is now in beta with gpt-6-luna, according to the company’s October 6 API changelog, which describes it as a way to turn text and images into typed answers up to 10 times faster than the Responses API [1]. A same-day OpenAI Developer Community announcement adds that the feature is available to all developers in public beta and is aimed at apps that need to choose a model, tool, or action in near real time [2].

That framing matters. The Decisions API is not presented as another general chatbot endpoint. It is a narrower interface for cases where the application already knows the set of acceptable outcomes: route the request to sales or support, choose the next UI action, classify an image into one of a few categories, or decide whether a live assistant should escalate to a heavier reasoning model. In exchange for giving the model a smaller answer space, developers get a response pattern that is easier to validate, easier to wire into product logic, and designed for latency-sensitive interaction.

The core workflow is simple: send input, define a question, provide possible answers, and use the returned choice in the application. OpenAI’s public beta post says Decisions supports three output types: predicates, choices, and scores [2]. Predicates estimate whether a statement is true; choices select from predefined options with confidence scores; scores evaluate an input against a numeric range [2]. In practice, that gives developers a compact decision layer between raw user or sensor input and deterministic software behavior.

Why GPT-6 Luna is the center of the launch

The launch ties Decisions directly to GPT-6 Luna. OpenAI’s changelog names gpt-6-luna as the model behind the beta and says the API turns multimodal inputs into typed answers 10 times faster than the Responses API [1]. The developer announcement repeats that the Decisions API is powered by GPT-6 Luna and can process text and image inputs [2].

The choice of Luna is telling. OpenAI is using a model with language and image understanding, but constraining its role. Instead of asking it to produce prose, reason through a long plan, or call tools in an open-ended sequence, the application asks it to decide among bounded answers. That is the architectural shift: intelligence remains probabilistic, but the interface is typed and limited.

This makes the API especially relevant for product teams building interactive systems. Many production AI features do not require a full paragraph of reasoning. They need a quick classification: Is this lead enterprise or self-serve? Is this image showing a blocked lane or a clear lane? Should the assistant reload the page, go back, do nothing, or hand the task to a stronger model? For those situations, a general response endpoint can be slower and harder to integrate than a decision endpoint.

The demonstrations show the intended pattern

Recent coverage of the launch demonstration describes several examples that keep the same structure: an input arrives, a constrained question is asked, and the application acts on the answer [3]. In one text example, unstructured order details are mapped into form fields, then similar inputs are classified for routing, such as sending support tickets or sales requests to the right team [3].

The most concrete latency signal came from a sales dashboard demonstration. Orply’s report says the interface showed 81 milliseconds of API processing for six decisions while classifying leads with attributes such as company type, team size, buying stage, requirements, timeline, and requested next step [3]. That is not a universal benchmark for every workload, but it illustrates the design goal: make model-assisted classification feel like part of the application’s own control loop rather than a separate conversational turn.

The visual examples are equally important. The demonstration used frames from a driving game in which the API chose among three lanes, with the application representing the valid actions as lane 1, lane 2, or lane 3 [3]. That example captures the advantage of bounded multimodal decisions: the model interprets an image, but the software still owns the action space.

Voice, agents, and robots: Decisions as a control layer

The API also fits into OpenAI’s broader move toward live assistants and agents. The launch material highlighted Decisions working alongside GPT-Live-1 and a Microduck robot, with the API helping add expression to voice conversations and guide the robot’s head movement from camera input [4]. Orply’s account of the same demo says GPT-Live-1 handled the spoken conversation while Decisions selected an expression from a set supplied by the application [3].

That separation of roles is significant. A voice model can maintain the conversation, while a decisions layer can choose a nonverbal reaction, a UI state, or a backend route. In a robot or animated character, the answer might not be a sentence at all; it might be “look left,” “smile,” “turn toward the apple,” or “take no action.” The API is therefore less about replacing agents than about giving agents and apps a fast, typed reflex.

OpenAI’s developer guidance for connecting voice to Decisions follows the same pattern: collect current app state and transcripts, send a prompt and available choices to POST /v1/decisions, read the chosen answer, and run the matching handler [2]. For more complex requests, the application can include an option that routes to a reasoning model instead of forcing a lightweight decision to do heavyweight work [2].

What developers should watch

The public beta gives developers a useful new primitive, but it also places responsibility on application design. The quality of a decision depends heavily on the question, the allowed answers, and the context sent with the request. A poorly designed choice set can make the system confidently choose among bad options. A well-designed one can make the model’s uncertainty manageable by including safe defaults such as “noop,” “unknown,” “escalate,” or “reason.”

The confidence-score aspect of choices also deserves careful treatment. Scores can help determine whether to automate, ask the user, or hand off to a human, but they should not be treated as universal truth. For high-impact workflows, teams will still need evaluation sets, latency measurements, fallbacks, logging, and human review thresholds. The appeal of Decisions is speed and structure, not a free pass around product safety.

The API’s biggest promise is architectural clarity. Modern AI apps increasingly combine multiple models, live audio, vision, tools, agents, and deterministic business logic. Decisions gives developers a small but important connector: a way to ask a multimodal model a bounded question and receive a typed result fast enough to drive the next step.

A small interface with large product implications

The Decisions API is important because it formalizes a pattern many developers were already building by hand. Instead of prompting a general model to “respond with one of these labels” and then parsing text, the API makes bounded decision-making a first-class endpoint. With public beta availability, GPT-6 Luna support, text and image inputs, and output modes for predicates, choices, and scores, OpenAI is positioning Decisions as the fast reflex layer for AI-powered software [1] [2].

If the beta performs consistently outside polished demos, its value will be clearest in places where milliseconds matter and open-ended answers are unnecessary: routing, moderation triage, visual state checks, voice UI actions, lead scoring, lightweight agent control, and escalation logic. The story is not that every task should become a tiny classification problem. It is that many product moments already are tiny classification problems, and OpenAI now has a dedicated API for them.

Sources from the last 72 hours

1. [1]Changelog | OpenAI APIOct 6, 2026, 2:00 AM
2. [2]Decisions API is now available in Public BetaOct 6, 2026, 10:53 PM
3. [3]A Small Set of Choices Makes Model Decisions Nearly 10 Times FasterOct 6, 2026, 2:00 AM
4. [4]Introducing the Decisions APIOct 7, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.