OpenAI's Decisions API: Faster AI Decisions for Real-Time Apps

On October 22, 2026, OpenAI announced the general availability of its Decisions API, a developer tool designed for applications that need frequent, small, clearly defined decisions. Instead of using a general-purpose model response for every classification or routing step, developers can specify the question, provide context, and define the kinds of answers the application is allowed to receive.

The API is built with GPT-6 Luna and is designed for bounded decisions such as classifying a support request, scoring a risk signal, selecting the next action in a workflow, or interpreting information from an image. OpenAI says these decision requests can run up to 10 times faster than using GPT-6 Luna through the Responses API for comparable bounded tasks. That is a vendor-reported claim, not a guarantee for every workload.

Official announcement: Build more responsive apps with the Decisions API.

What Is the Decisions API?

Most AI integrations ask a model to generate an open-ended response. That approach is useful for explanations, writing, and flexible conversations, but many application tasks need a narrower result. A support system may only need to know whether a message belongs to billing, technical support, shipping, or another category.

The Decisions API is designed around that kind of question. Developers define the decision they need, supply the relevant input, and describe the allowed outcomes. The API returns a structured answer that application code can evaluate and use in its workflow.

The Three Decision Types

1. Predicate

Use a predicate when an application needs an estimated probability that a statement is true. Examples include checking whether a product image appears to show visible damage or whether a message contains a particular kind of request.

2. Choice

Use a choice when the possible answers are known in advance. A support application could select billing, technical, shipping, or other. The response includes the selected value along with probabilities and a confidence field that can support a review policy.

3. Score

Use a score when an input needs to be assessed against an ordered rubric. For example, an issue could be classified as cosmetic, recoverable with a workaround, or fully blocking. Define each level clearly so that neighbouring scores have distinct meanings.

These answer types are designed to make the model's output easier to integrate into an application. They do not remove the need to test whether the decision is reliable for your users and data.

Why This Matters for Developers

A small decision can sit on the critical path of a much larger workflow. Reducing the latency of repeated classification or routing calls can make the whole product feel more responsive, especially when the workflow happens during a live conversation or before an agent takes its next step.

OpenAI says the Decisions API supports text, images, and structured JSON inputs. Developers can ask up to 400 independent questions in one request, reuse context through prompt caching, and integrate decisions with GPT-Live for real-time voice experiences. Regional processing is available in the United States and European Union, and Zero Data Retention is available to eligible customers. Check the official documentation for current availability and eligibility.

Potential use cases include:

Customer support: Route incoming messages to the appropriate team.
Agent safety: Score a proposed command or action before it is allowed to run.
E-commerce: Classify product images or identify cases that need manual review.
Finance operations: Categorise transaction descriptions for downstream processing.
Voice applications: Choose whether to clarify, continue, or escalate during a live conversation.
Quality workflows: Score an issue against a documented severity rubric.

A Node.js Example

The official OpenAI JavaScript SDK exposes the endpoint through client.decisions.create. Install the SDK first:

npm install openai

Set OPENAI_API_KEY in the server environment, then use a small support-ticket classifier like this:

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
});

const result = await client.decisions.create({
  model: "gpt-6-luna",
  input: "I was charged twice for my order.",
  questions: [
    {
      type: "choice",
      name: "department",
      instructions: "Which department should handle this complaint?",
      choices: [
        { value: "billing", description: "Payments, invoices, and refunds." },
        { value: "technical", description: "Problems using the product." },
        { value: "shipping", description: "Delivery and tracking." },
        { value: "other", description: "Requests outside these categories." },
      ],
    },
  ],
});

const answer = result.answers[0];

if (answer?.type === "choice") {
  console.log({
    department: answer.choice,
    confidence: answer.confidence,
  });
} else {
  console.log("Send this request for manual review.");
}

This example follows the request structure in the official documentation. It illustrates integration rather than reporting a benchmark or a tested production deployment. Keep the API key on the server and never expose it in a browser bundle or a NEXT_PUBLIC_ environment variable.

Decisions API vs. a General-Purpose Response

Use the Decisions API when the question and acceptable answers can be defined ahead of time, and the result will feed into a specific application step. This includes classification, scoring, routing, or a constrained choice during a live interaction.

Use a general-purpose response workflow when you need a longer explanation, free-form content, a multi-turn conversation, or broader tool orchestration. Choosing a specialised endpoint does not automatically make an application better. The right option depends on the shape of the task and the evidence from your own evaluation set.

Pricing and Availability

At the time of writing, OpenAI lists Decisions API input pricing at $0.10 per million uncached input tokens and $0.01 per million cached input tokens. The announcement says there are no output-token charges for this endpoint. Cache writes and other processing options may have separate rates, so verify the current pricing and availability guide before estimating production costs.

The endpoint is generally available with GPT-6 Luna. Availability, regional processing, and retention options can depend on account eligibility and the selected configuration.

Important Engineering Considerations

A structured answer is not the same as a guaranteed correct decision. Before connecting model output to production actions, build a representative evaluation set, measure false positives and false negatives, and choose thresholds based on the cost of being wrong.

Recommended practices include:

Define clear outcomes: Make choices mutually distinct and provide specific instructions for each option.
Use confidence thoughtfully: Low-confidence or ambiguous cases should be sent to a human or a safer fallback path.
Keep permissions separate: A model classification should not by itself grant access to sensitive data or authorise a destructive action.
Evaluate with real examples: Test edge cases, unusual inputs, and changes in user behaviour before and after release.
Monitor the production workflow: Log the decision, its outcome, and relevant latency or failure metrics while respecting privacy requirements.
Recheck pricing and policies: API capabilities and account options can change, so confirm current documentation before deployment.

Final Thoughts

The Decisions API reflects a broader direction in application development: AI systems are being designed not only to generate content, but also to make focused decisions inside larger software workflows. Its value will depend on whether teams can combine low latency with clear answer boundaries, careful evaluation, and reliable fallbacks.

For developers, the practical starting point is small. Pick one repeated classification or routing task, write down the allowed outcomes, compare the API against the current implementation, and measure both speed and accuracy before expanding its role.

Official Resources

Information checked against OpenAI's announcement and developer documentation on October 25, 2026. Pricing and product capabilities may change.