OpenAI Decisions API: Classify, Route, Score
Turn yes/no checks, routing, and severity ratings into typed answers with OpenAI's Decisions API (gpt-6-luna), plus the thresholds and human review they need.
Most AI tools give you a paragraph. The OpenAI Decisions API gives you a number, a label, or a rating, and nothing else. Ask “Does this photo show visible damage?” and you get a probability. Ask “Which department should handle this complaint?” and you get one of your own options with a probability for each. Ask “How severe is this bug?” and you get a score against levels you wrote. That makes it the right tool for the boring, high-volume decisions that clog support queues, intake forms, and review backlogs.
It also makes it easy to misuse. A probability of 0.92 feels like “92% accurate.” It is not. It is the model’s estimate, and whether that estimate matches reality on your data is something you have to measure. So this course spends as much time on thresholds, calibration checks, refusal handling, and the human-review lane as it does on the three question types. By Lesson 2 you’ll have written a real predicate question for a decision from your own work and read its answer. By Lesson 5 you’ll know how to compare candidate thresholds using the cost of each kind of mistake and labeled examples. By the capstone you’ll have a one-page Decision Design Brief a developer can build from and a manager can approve.
Every product fact comes from OpenAI’s official Decisions guide, model page, pricing page, and changelog, all read on October 13, 2026, one week into the public beta. Where the only evidence is a developer-forum report, the lesson says so. And because the API is still in beta with a single model, each lesson that quotes a field name, price, or limit tells you exactly what to re-check before you rely on it.
What You'll Learn
- Explain what the Decisions API returns, which model runs it, and how it differs from Structured Outputs and function calling
- Use the exact request shape to write a predicate question with observable criteria and read its probability
- Design choice and score questions with distinct options, a fallback, and ordered levels
- Evaluate a threshold using the cost of false positives and false negatives, and explain why a probability is not an accuracy guarantee
- Design refusal handling and a human-review path for every decision you automate
- Create a Decision Design Brief with a beta go-live checklist for one real decision at work
After This Course, You Can
What You'll Build
Course Syllabus
Prerequisites
- Comfort reading a short block of JSON (we explain every field); no coding required
- To run the optional examples: an OpenAI API account with billing enabled and an API key
- A real repetitive yes/no, sorting, or rating decision from your work to use as your running example
Who Is This For?
- Support, operations, and compliance leads who want AI to sort, flag, and prioritize work without losing control
- Product managers and analysts who write specs for developers and need to get the decision logic right
- Developers new to the Decisions API who want the design and review side, not just the request shape
- Anyone who has been handed a 'confidence score' and wants to know what it actually means
Frequently Asked Questions
What is the OpenAI Decisions API?
A dedicated endpoint (POST /v1/decisions) that evaluates text, images, or both and returns typed answers instead of free text: a probability that a condition is true (predicate), one option from a fixed set (choice), or a rating against ordered levels (score). OpenAI positions it for classifying content, routing requests, and prioritizing work. As of October 13, 2026 it is in public beta and runs only on gpt-6-luna.
Do I need to be a developer?
No. Six of the eight lessons are about designing good questions, choosing thresholds, and setting up human review, which is where most real-world failures happen. The exact request and response shapes are shown as vendor-verified JSON you can read without coding. If you do code, every example mirrors OpenAI's own documentation.
How is this different from Structured Outputs or function calling?
Structured Outputs makes a model generate an object that follows your JSON schema (extracted fields, a written explanation). Function calling makes a model request a tool call with arguments. Decisions returns one of three fixed answer types with probabilities, and nothing else. Lesson 1 gives you a decision table for all three.
Is a 0.92 probability the same as 92% accuracy?
No, and this course keeps saying so. The number is the model's estimate, not a measured hit rate. Lesson 5 shows how to set a threshold from labeled examples and the cost of each kind of error, and how to check whether the model's numbers match reality on your data.
What does it cost?
Per OpenAI's guide on October 13, 2026: $0.10 per 1 million input tokens on gpt-6-luna, with no output, cache-read, or cache-write charges on this endpoint. Regional processing premiums and long-context multipliers still apply. Beta pricing can change, so the capstone checklist tells you to re-read the pricing page before launch.
