OpenAI Decisions API: Typed AI Answers at $0.10 per Million Tokens
The OpenAI Decisions API hit public beta on October 6: one endpoint that returns probabilities, choices or scores on GPT-6 Luna, billed only for input at $0.10 per 1M tokens.
- OpenAI announced the public beta on October 6 in its developer forum, saying the endpoint makes decisions up to 10x faster than GPT-6 Luna through the Responses API.
- Three answer types: predicates (probability a statement is true), choices (one option from a list, with confidence) and scores (a rating on ordered levels).
- Only input tokens are billed. No charges for output, cache reads or cache writes. General availability is expected "in the coming weeks," per Unite.AI's write-up of the docs.
in this block
The OpenAI Decisions API is now open to every developer in public beta. It is a dedicated endpoint, POST /v1/decisions, that takes text or images plus a list of questions you define and returns typed answers: probabilities, picks from a fixed menu or scores. It runs on GPT-6 Luna and costs $0.10 per million input tokens, with output tokens free.
What actually happened
The endpoint first showed up at DevDay in limited preview. On October 6 OpenAI flipped it to public beta for all developers. The pitch is narrow on purpose: let your app pick the right model, tool or action in near real time instead of asking a chat model to write a paragraph and then parsing it.
Each request has three parts. The model field, which for now only accepts gpt-6-luna. The input, either a text string or user messages with text and images. And a questions array, where every question has a name, a type, instructions and any allowed choices or score levels. The response returns an answers array keyed to those names.
GPT-6 Luna itself is not new. Unite.AI notes OpenAI released it alongside GPT-6 Sol on September 22. We covered Sol's debut in GPT-6 Sol.
How the three answer types work
A predicate checks a condition, like "is there a crack, tear or dent in this product photo?" It returns a probability from 0 to 1. In the docs example, a damaged-product photo gets 0.92, and your app decides the threshold for flagging it.
A choice picks one value from options you supply, each with a short description. It returns the pick, a probability for every option and a separate confidence. The routing example sends a double-charge complaint to billing instead of technical or shipping. OpenAI recommends adding an "other" fallback.
A score rates the input against ordered levels, like issue severity. The score is a probability-weighted average of level indices, so it can land between levels. In the worked example, probabilities of 0.1, 0.7 and 0.2 across three levels give a score of 1.1.
There are guardrails. Images must be sent inline as base64 data URLs, not hosted links or file IDs. Dependent decisions need separate requests. The API supports Zero Data Retention and HIPAA for eligible customers, with US and European data residency.
Early vibes from builders
The forum thread lit up fast. One developer joked about benchmarking it against Jev, the decision model we profiled in JEV-27B decision model. Another posted quick tests on a few hundred word-game and conversational questions and concluded Jev was faster, cheaper and more accurate on nuanced judgment calls, with Luna roughly even on simple yes/no checks.
Treat that as one user's small test, not a benchmark. Others pointed out a real gap in Luna's favor: image understanding. If you need to check whether a picture breaks site rules or contains source code, the OpenAI Decisions API handles that natively.
Classifiers are back, but now they come with a probability attached. Big if true for every agent router out there.
What to do as a reader
If you build agents or moderation pipelines, this is worth a weekend test. Take a few hundred labeled examples from your real traffic, run them through the endpoint and set thresholds based on the cost of false positives versus false negatives. That is literally OpenAI's own guidance.
Do not migrate production on day one. It is a beta, only one model is supported, and pricing still has regional and long-context multipliers. Compare it against whatever you use now, including small open models, on speed, cost and calibration.
If you are just here for the hype, the takeaway is simple. The OpenAI Decisions API is a cheap, fast "yes, no or which one" button for AI agents. Those buttons are what turn chatbots into products that actually route tickets, approve refunds or pick the next tool. You can read the official announcement and the docs before you spend a cent.
Not financial advice, not engineering gospel either. Test it on your own data, ser, and keep a human in the loop for anything that touches money.
Not financial advice. DYOR, ser.
