Tech • AI • Robotics • Game
Decisions API Full Tutorial for Beginners: How to Use & Set Up Decisions API
8/10TL;DR
OpenAI’s Decisions API is designed for software workflows that need probabilities for routing, prioritization, and classification rather than generated text, and it can evaluate both text and images.
KEY POINTS
A shift from generation to decisioning
Most business software tasks do not need a paragraph-length answer. They need a direct signal such as yes or no, a priority level, or a category. OpenAI’s Decisions API is built for that use case, returning probabilities that applications can act on immediately instead of text that a person or another system must parse.
Faster and cheaper for simple workflows
Unlike the Responses API, which generates text, the Decisions API skips generation and focuses on evaluation. That makes it about 10 times faster, while also removing charges for output tokens. Pricing is listed at $0.10 per million input tokens, making short classification and routing tasks inexpensive.
One model, three decision types
The API currently runs on a single model, GPT-6 Luna, and supports three formats. Predicate checks how likely a statement is true, score rates an input across ordered levels, and choice selects the best option from a predefined list. Together, those formats cover common automation tasks such as spam checks, urgency triage, ticket routing, and image labeling.
Predicate handles binary-style checks
Predicate is intended for one clear condition at a time, written as a statement rather than a question. In a bakery support example, a message saying an order arrived an hour late and the box was soaked by rain strongly matched a delivery-problem predicate, while a note that a cake was “delicious, just a little too sweet” did not. The output is a probability from 0% to 100%, and businesses still need to set an action threshold such as 80% before routing or flagging messages.
Score ranks urgency instead of simply detecting it
Score is meant for ordered judgments such as urgency, severity, or importance. In one example, three levels were defined: can wait, reply today, and respond now. A message about a birthday cake pickup in 2 hours that had been unexpectedly canceled strongly pointed to the highest urgency tier, while a question about gluten-free options for next week landed lower.
Scores can fall between levels
The final score is not always a whole number because it is calculated as a probability-weighted average of the defined levels. That means an application can see not only the likely category but also how strongly the model leans in that direction. The playground also shows a confidence value, which can help decide when human review is still necessary.
Choice supports routing into distinct categories
Choice is aimed at sorting requests into one option from a non-ordered list. A bakery message asking to change a cookie order from one dozen to two dozen would likely be routed to change an order, while a question about delivery to the east side of town would map to general question. Keeping categories distinct matters, because overlapping options can spread probabilities and make routing less reliable.
Image classification uses the same framework
All three decision types also support image inputs, and text and images can be combined in one request. In a product-sorting example, five options were defined: bread, pastry, cake, cookie, and something else. A croissant should rank as pastry, a sourdough loaf as bread, and an unrelated image such as pizza should fall into the fallback category.
No-code testing is available in the playground
The API can be tried directly in the Decisions Playground at platform.openai.com/decisions, with no project setup or coding required. The interface includes prompt fields, decision settings, templates, input handling, and visible probability outputs. That makes it suitable for testing workflow logic before developers integrate it into production systems.
CONCLUSION
The main appeal of the Decisions API is not content generation but operational automation. By returning structured probabilities for text and images, it gives businesses a practical way to route, rank, and classify work at speed and low cost.
Ask a question