Skip to content
Main Site News Console

Build More Responsive Apps with the Decisions API

· OpenAI Translated
OpenAILLM

Building more responsive applications with the Decisions API | OpenAI

October 22, 2026

Fast, focused, and cost-effective decision-making for any application.

Applications and agents make countless subtle decisions throughout a workflow: how to classify an input, where to route a request, or what action to take next. Individually, each decision is simple, but at scale, latency and costs compound quickly.

Today, we are announcing the general availability (GA) of the Decisions API. Built on GPT‑6 Luna, the Decisions API focuses Luna’s intelligence on questions with well-defined answers, helping applications and agents make near real-time decisions from text, images, and structured inputs—up to 10x faster than calling GPT‑6 Luna through the Responses API.

Since the public beta, we have added native GPT‑Live integration, prompt caching for reusing context, and inference data residency support across the US and EU. Developers can also pass native JSON inputs, include images in question instructions, and ask up to 400 questions per request—doubling the previous limit. In addition, we are continuously optimizing latency and have expanded compute capacity across multiple regions in Asia.

How teams are using the Decisions API

Early adoption of the Decisions API demonstrates its versatility across a wide range of use cases.

Inside OpenAI, teams are applying the Decisions API across product and operational workflows. The dots team is using it to reduce response times and boost interactivity; the personal finance team uses it to categorize transaction descriptions; and the communications team uses Decisions to classify press coverage by topic and assess whether it reinforces or undermines key narratives—helping highlight what requires attention.

These early examples demonstrate how fast, structured decisions turn incoming information into effective action across diverse use cases. Customers are using the Decisions API just as broadly, from accurately attributing AI costs to building more engaging language learning experiences.

“With Box Extract and the Decisions API, we can classify business documents and route them to the right next step. In our evaluations across 30 Certificate of Insurance cases, the Decisions API routed decisions ~5x faster with the same accuracy as the Responses API—automating routine work while routing lower-confidence cases to human reviewers.”

—Biswaroop Palit, VP of Product Management, Platform and Integrations at Box

“The Decisions API helps us evaluate the risk of agent-generated rshell commands before executing them to autonomously remediate incidents. Compared to existing frontier models, it reduced p95 latency by 40x while maintaining comparable accuracy—making it viable to place Decisions directly in the hot path of remediation actions.”

—AJ Stuyvenberg, Staff Engineer at Datadog

“Compared to GPT‑6 Luna, the Decisions API enabled us to attribute AI costs to actual tasks 4x faster and at 30% lower cost, without sacrificing accuracy.”

—René Sultan, Research Engineer at Ramp

“At Sierra, we use the Decisions API to select the next action in our real-time customer conversation orchestration flows. With p95 latency under 280ms, it is roughly 3x faster than our previous approach and cuts cost per invocation by 90%, all with zero compromise to customer outcomes.”

—Ajeet Grewal, Engineer at Sierra

“For language learners, immediacy is critical. In our testing, the Decisions API integrated with GPT‑Live reduced next-action response times by over 90% (from 1.4s to 0.13s), creating a far more immersive, interactive learning experience.”

—Andrew Hsu, CTO at Speak

Choosing the right decision type for your application

Using the Decisions API is straightforward: developers define a question and its allowed set of answers, then provide context in the form of text, images, or native JSON. The API returns one of three supported answer types—probability, choice, or score—which applications can use to determine the next action. These diverse answer types make Decisions adaptable to a broad spectrum of scenarios.

The Decisions API supports three answer types:

TypeBest forKey result
predicateAssessing the probability that a condition is true, such as whether a product has visible damage.probability: an estimated value between 0 and 1 indicating whether the condition is true.
choiceSelecting from a set of predefined options, such as whether an agent should proceed, retry, or ask for clarification.choice: one of the provided candidate values. The response also includes probability and confidence.
scoreRating an input against an ordered scale in a rubric, such as the severity of damage.score: the probability-weighted average of the rubric level indices.

Building real-time voice experiences with the Decisions API and GPT‑Live

With GPT‑Live, you can bring natural, expressive conversational capabilities to voice applications, complete with smooth interruption handling to keep conversations flowing naturally. The Decisions API is now natively integrated with GPT‑Live, helping applications make fast, precise decisions during live conversations. For example, a customer support agent can decide whether to ask a clarifying question, route the request, or escalate the ticket; an avatar can adjust its facial expressions while a user speaks, responding with a wink or a smile based on what the user says to create a more interactive experience.

The following example demonstrates how to use transcript updates to select an avatar’s reaction. You can add this configuration to session.triggers in an existing GPT‑Live session:

JSON

{
"type": "decisions",
"event_type": "session.input_transcript.delta",
"decisions": {
"questions": [{
"type": "choice",
"name": "reaction",
"instructions": "Choose an appropriate reaction for the ava