---
title: "OpenAI Decisions API: Pricing, Decision Types, Use Cases"
url: "https://goremy.ai/blog/articles/openai-decisions-api-pricing-how-it-works"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-17T17:40:10.170Z"
updated: "2026-11-17T17:40:10.170Z"
topic: AI Pricing & Costs
description: "OpenAI's Decisions API explained: predicate, choice and score decisions, the $0.10 per million input token price with no output charge, and when to use it."
---

# OpenAI Decisions API: Pricing, Decision Types, Use Cases

> OpenAI's Decisions API explained: predicate, choice and score decisions, the $0.10 per million input token price with no output charge, and when to use it.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-17

## What is the OpenAI Decisions API?

The Decisions API is an OpenAI interface built for quick judgments instead of written answers. You give it an input, such as a photo, an email or a bug report, and it returns a decision your software can act on right away: a probability, a pick from a list, or a score. It is meant for cases where you do not need a chat-style model to write out a long response and only need a quick decision. OpenAI priced it at $0.10 per million uncached input tokens with no output token charges.

## TL;DR

- The Decisions API returns **short, structured judgments** (a probability, a choice or a score) instead of paragraphs of text.
- There are **three decision types**: predicate (yes/no with a probability), choice (pick from a list) and score (rate on a scale).
- OpenAI says decisions can be **up to 10 times faster** than using the Responses API with GPT-6 Luna. That is a vendor claim, so test it on your own workload.
- Pricing is **$0.10 per million uncached input tokens**, and **output tokens are not charged**.
- Good fits are **triage and routing jobs**: sorting email, flagging damaged returns, rating issue severity, or helping an AI agent decide whether to retry or ask for help.
- It is not a replacement for writing, summarizing or explaining. Use the Responses API when you need actual prose.

## Why does a decisions-only API exist?

Most people think of AI as something that writes. You ask, and it produces a paragraph, a draft or a block of code. But a lot of real work inside a company is not writing. It is deciding. Is this email urgent? Does this package look damaged? Which team should get this ticket?

If you use a general chat-style model for those jobs, you pay for and wait on a lot of text you do not need. The model may produce a sentence of reasoning when your software only needed the word "urgent."

A decisions-focused API removes that overhead. The output is small and predictable, so your app can use it directly. A person does not have to read it, and your code does not have to dig an answer out of a paragraph.

## What are the three decision types?

The Decisions API supports three kinds of decision.

| Decision type | What it returns | Example |
|---|---|---|
| **Predicate** | A probability that something is true | How likely is it that this returned package has a crack or dent? |
| **Choice** | One option picked from a list you provide | Which folder does this email belong in? |
| **Score** | A rating on a scale | How serious is this technical issue? |

### Predicate: how likely is this true?

A predicate is a yes/no question answered with a probability instead of a flat yes or no. The example OpenAI has shown is a returns app. You send in a photo of a returned package, and the API returns the probability that the item is damaged. Your app can then set its own rule, such as "flag anything above a certain probability for human review."

The probability is the useful part. A hard yes or no hides how confident the model is. A number lets you decide how cautious to be. A store worried about fraud might review at a low threshold. One that wants to cut manual work might set it higher.

### Choice: which option fits best?

A choice decision picks from a list you define. Sorting an inbox is the obvious case. You supply the categories, say "billing," "scheduling," "sales lead" and "ignore," and the API picks one for each message.

### Score: how much, on a scale?

A score decision rates something on a scale. Scoring how serious a technical issue is falls here. Scores are handy for ranking a queue, so the most severe items rise to the top.

## How fast is it compared with the Responses API?

OpenAI's claim is that decisions can be up to 10 times faster than using the Responses API in GPT-6 Luna. Two details matter here.

First, "up to" means the best case. Your results will depend on the input, the task and how you set things up.

Second, the comparison is against a general-purpose route doing the same job. It does not say Decisions is faster at everything. The speed gain comes from the output being tiny and structured.

Speed matters most in two situations. One is when a decision sits in the middle of a user-facing flow, like flagging a return while the customer waits. The other is when an AI agent makes many small decisions in a row and each delay adds up. Before you build anything around the 10x number, run a small test with your own data and measure the actual difference.

## How much does the Decisions API cost?

The pricing OpenAI announced has two parts:

- **$0.10 per million uncached input tokens.**
- **No output token charges.**

A token is a small chunk of text, roughly a short word or part of a word. "Uncached" means input the system has not already seen and stored. The announcement gave only the uncached rate, so check OpenAI's pricing page for how cached input is billed.

Here is rough arithmetic to make the price concrete. This is an illustration, not an OpenAI figure. Suppose each email you classify is about 500 tokens long. A thousand emails is about 500,000 tokens, which at $0.10 per million comes to about five cents. Images are counted differently from text, so a damage-check workload needs its own estimate. Run a small batch and read the usage numbers.

Because output is free, the price is predictable. You pay for what you send in, and the size of the answer does not change the bill. With a regular chat model, long answers cost real money.

## When should I use it instead of the Responses API?

A simple rule: if the result is something your software acts on, consider Decisions. If the result is something a person reads, use the Responses API.

**Use the Decisions API when:**
- You are sorting, routing or flagging items in volume.
- You need a number or a label, not an explanation.
- Speed and cost per item matter because you run it thousands of times.
- An AI agent needs to choose its next step, such as keep going, retry, or ask a human for help.

**Use the Responses API when:**
- You need written output: a draft, a summary, an explanation.
- The task has several steps of reasoning that you want spelled out.
- You need the model to explain why it reached a conclusion.

A common pattern is to use both. Decisions handles the first pass over a large pile of items, and the Responses API writes up only the small set that needs attention.

## What can you actually use it for?

The uses OpenAI has pointed to are practical ones:

1. **Email triage.** Choose a folder or priority for each incoming message.
2. **Severity scoring.** Rate how serious a bug report or support issue is so engineers see the worst first.
3. **Return inspection.** Estimate the probability that a returned item is damaged from a photo, then route likely cases to a person.
4. **Agent control.** Let an agent decide whether to continue, retry a failed step, or stop and ask for help.

Outside those, any repetitive "look at this and decide" task has the same shape: expense receipts that look unusual, resumes that match a role, support tickets that mention cancellation. These are our own examples of the pattern, not announced features, so verify accuracy before you rely on them.

## What does this mean for you?

**If you work in operations, support or finance,** your team probably has a queue someone sorts by hand. Look for the one where the decision takes a person a few seconds and happens hundreds of times a week. That is the candidate. Start by writing down the categories or the scale a human uses today.

**If you manage people,** treat the probability output as a staffing tool. You can decide how much goes to human review by setting a threshold, and you can change it as you learn how accurate the system is.

**If you are not technical,** you do not need to call the API yourself. But you can now ask your engineering or IT team a sharper question: "Which of our manual sorting jobs only needs a label or a score?"

**What to do first:**
1. Pick one narrow task with clear categories.
2. Collect 100 past examples with the correct answer a human gave.
3. Run them through and compare. Check where it is wrong, not just how often.
4. Keep a person in the loop for high-stakes calls, such as refusing a refund or rejecting a candidate.

One caution: a probability is not a guarantee. A model that says 90% likely damaged will still be wrong some of the time. Plan for that.

## Could you build a small decision tool of your own?

If a Decisions-style workflow appeals to you but you do not want another subscription, the missing piece is usually the app around the AI call: a place where items arrive, get flagged, and wait for a human. With [Remy](https://goremy.ai), you describe that tool in plain English, such as a returns review queue where staff upload photos, see a damage flag, and approve or reject. Remy builds it as a full-stack app with a database, sign-in and a live URL, and it can call AI models for classification. Check which models and endpoints a build can use before you plan around one specific OpenAI feature.

## Frequently Asked Questions

### What does the OpenAI Decisions API cost?

OpenAI priced it at $0.10 per million uncached input tokens, with no charge for output tokens. The announcement did not give a cached-input rate, so check OpenAI's pricing page for the full picture before budgeting.

### Is the Decisions API really 10 times faster?

OpenAI says decisions can be up to 10 times faster than using the Responses API in GPT-6 Luna. That is a best-case claim from the company, and your results will vary by task and input. Measure it on a sample of your own data.

### What is the difference between predicate, choice and score?

A predicate returns the probability that a statement is true, like "this package is damaged." A choice picks one option from a list you provide. A score rates something on a scale, such as issue severity.

### Can the Decisions API write text for me?

No, that is not its purpose. It is built to return short, structured judgments your software can act on. For drafts, summaries or explanations, use the Responses API.

### Do I need to be a developer to benefit from it?

You need a developer to connect it to your systems, since it is an API. But anyone who owns a sorting or triage process can identify the right tasks and define the categories, which is the hardest part of making it work.
