DecisionsApi · Focused questions. Predefined answers.

OpenAI Decisions APILuna model guide

OpenAI's Decisions API uses Luna's capabilities to answer your questions by choosing from a finite set of predefined answers. Its announcement describes text or image context for classification, request routing, and agent decisions. Explore the workflow here, then test the models supported by this independent platform.

Test a decision workflow See an integration example

Read OpenAI's official DevDay announcement

Define the expected answerCompare the same examplesInspect uncertain cases

This platform's output example

Static illustration

Customer message

“I was charged twice this morning. Please refund the duplicate payment today.”

One message → three useful decisions

Choice

Billing

Choose the support queue

Yes / no

96%

Signal for an urgent request

Score

2.8 / 3

Urgency on a defined 0–3 scale

Static illustration of this platform's decision format. These values are not a Luna response, a live result, or an accuracy benchmark.

01 / Try the workflow

Build a baseline with supported models

Use your own text, define the answers you expect, and inspect the request and response. The workbench runs the models listed in its selector; this page does not provide a Luna demo.

Check the selected model and the workbench's current usage terms before running a request.

Workbench

Example library

What do you want to decide?

Choose a scenario to load editable input and rules. Loading an example uses no credits.

Examples load into the current draft only; no configurations or run history are saved.

ModelInput format

State

Give the selected model the context shared by every question.

Text

Input state

All questions share this context and are answered independently.

Questions

Define up to eight typed questions for the same state.

3 / 8

3. Run decision

Run the decision request.

Result preview

Subject: Charged twice again!! Hi — this is the SECOND month in a row I've been billed twice for the Pro plan. I already emailed last month and nobody replied. I run my whole business on this. If it's not refunded today I'm cancelling and disputing the charge with my bank.

↓One input → multiple structured decisions

Click “Generate decisions” to see typed output with probabilities.

Results are returned in a structured format so you can connect them to your application logic.

02 / Plan the evaluation

Turn model research into a repeatable test.

A useful comparison starts with a fixed task, clear labels, and representative inputs. Use the same criteria when assessing any model your account can access.

Collect realistic inputs

Choose messages and records from the workflow, including short requests, missing context, and ambiguous examples.

state = ticket.text

Define a good answer

Write the allowed labels, scoring criteria, or yes-and-no rules before comparing model responses.

questions = { intent, urgent }

Review each result

Compare outputs against reviewed answers. Record errors, response time, and actual usage for each tested model.

answer = result.answers.route

Set the action boundary

Decide which results can trigger an action and which require more context or human review before deployment.

if result.urgent: escalate()

03 / Choose a useful task

Evaluate decisions your product already makes

OpenAI's announcement highlights classification, request routing, and an agent's next action. Use these related scenarios to prepare a practical test set for your OpenAI Decisions API Luna model evaluation.

Route customer requests

Define billing, technical, account, and other queues. Test messages that mention several issues and measure how often the intended queue is selected.

Choose an agent's next step

Compare record lookup, tool use, text generation, and human review. Include cases where a requested action lacks sufficient information or permission.

Check a claim against evidence

Pair a claim with a source passage and label it supported, contradicted, or undetermined. Keep missing evidence separate from a false claim.

04 / Choose the output

Three decision formats to test on this platform

A clearly defined output makes results easier to compare and connect to code. Use the formats supported by your selected model; confirm any OpenAI model's output contract in its official documentation.

Choice: one defined label

Set a small list of distinct options and describe where their boundaries fall. Include an other or review option when a case may not fit.

Score: an explicit scale

Describe each level of urgency, relevance, or completeness. Compare disagreements by level so a small difference is distinguishable from a serious error.

Noul: a focused yes–no check

Define what yes and no mean, then inspect the returned probability on held-out examples before choosing an automation threshold.

From evaluation to implementation

Inspect this platform's integration pattern

This server-side example calls this platform's /v1/systemone endpoint with typesafe/jev-1.13. Use it to understand state, questions, authentication, and typed answers. For an OpenAI integration, obtain the model ID and request format from the official documentation for your accessible model.

Read this platform's API docs

Server-side REST request

Typed decision answers

This platform's API key

POST /v1/systemone · this platform

JavaScript example · typesafe/jev-1.13 · not a Luna API request

POST /v1/systemone

const response = await fetch(
  "https://decisionapi.net/v1/systemone",
  {
    method: "POST",
    headers: {
      "Authorization": `Bearer ${DECISIONS_API_KEY}`,
      "Content-Type": "application/json"
    },
    body: JSON.stringify({
      model: "typesafe/jev-1.13",
      state: ticket,
      questions: { route: { type: "choice", criteria } }
    })
  }
);

05 / Understand the service

Start building with Decisions API

Testing credits included

New users receive credits for testing, so you can create an account, load an example, review the budget, and run your first questions.

One wallet for every live call

Use the same credit balance across the playground, API, and evaluation workflows. Credits from purchased packs do not expire.

Usage you can inspect

Review request history and credit activity, and use request identifiers to connect API calls with their usage records.

Plans for this independent platform

These plans cover this platform's services and supported models. Choose a plan for your testing and integration needs; the amounts shown are not OpenAI Luna pricing.

A purchase here does not grant OpenAI model access. Check official OpenAI documentation and your OpenAI account for model availability, eligibility, and pricing.

Important: Decisions API is currently in limited testing

Decisions API is not fully open yet and is currently available only for a small-scale test. Access will be enabled in payment order, and we’ll notify you by email. In the meantime, you can use Jev AI and other Decision APIs.

Starter

$10

Validate one real workflow, from the playground to your first API call

100,000 credits, no expiry
1 workspace
3 concurrent requests
Standard speed
Choice, score, and noul questions
Typed decision output
Probability and confidence results
Online playground
API key management
Email support

Recommended

Pro

$100

Connect classification, routing, and safety decisions to a production product

1,000,000 credits, no expiry
Unlimited workspaces
10 concurrent requests
Fast lane
Choice, score, and noul questions
Typed decision output
Parallel questions per request
API access
Usage and request history
Priority support

Enterprise

$1,000

For multiple workspaces, team collaboration, and custom production integrations, with 10% extra credits

11,000,000 credits (10% extra included)
Everything in Pro
Unlimited concurrency
Dedicated fast lane
Team workspaces and collaboration
Custom integration support
Security and permission guidance
Dedicated support
Priority production processing
Product roadmap feedback

Compare on your own data

Build an evidence-based model shortlist

Start with the models available here, then apply the same reviewed test set to other models you can access. Track decision quality, response time, and actual cost without treating different providers' probability values as interchangeable.

Quality on reviewed examplesMeasured response timeActual cost per decision

Explore available model comparisons

Jev

Establish a baseline for choice, score, and yes–no tasks

Laya

Compare results using the same inputs and decision criteria

Kev, Solar, and Span

Add available models and inspect their supported question types

Recent blog posts

What we are learning

Product notes, technical details, and practical ways to use Jev.

View all posts

AI APIs

Decisions API vs Jev: Differences & Integration Guide

Compare Decisions API vs Jev: platform and model differences, typed outputs, API integration, cost evaluation, and a practical checklist for choosing your stack.

Oct 3, 2026 · 11 min read

AI Models

OpenAI Decisions API Luna Model: A Practical Guide

Understand the OpenAI Decisions API Luna model, preview status, bounded decisions, routing contracts, evaluation, and a practical path to production.

Oct 3, 2026 · 12 min read

Developer Guide

Decision API OpenAI: Integration Guide and Practical Examples

Build a decision API OpenAI workflow with typed outputs, a REST example, evaluation metrics, and production controls. Understand the different API contracts.