OpenAI opens its Decisions API, and output is free

The classifier you built out of a chat model and a JSON schema now has its own endpoint.

OpenAI released the Decisions API in beta on October 6, according to its API changelog. It runs one model, gpt-6-luna, and returns typed answers instead of prose.

A question in, a probability out

The Decisions API guide describes three kinds of question:

• A predicate returns the probability, from 0 to 1, that a condition is true.
• A choice picks one option from values you supply and returns probabilities and a confidence.
• A score rates the input against ordered levels you define and returns a probability-weighted average score with a confidence.

It takes text, images or both. Images must be sent inline as base64 data URLs. Hosted image links and file IDs are not supported. Requests go to POST /v1/decisions.

OpenAI's examples are plain work: checking an image for visible damage, routing a customer complaint to the right department, and scoring an issue against severity levels.

The price is the part to read twice.

$0.10 per million input tokens. Output tokens cost nothing.

Cache reads and cache writes are also free, per the guide, though regional processing premiums and long-context multipliers still apply. OpenAI says the API answers "10x faster than the Responses API." That is OpenAI's claim, and your traffic is the test.

OpenAI says it expects general availability "in the coming weeks." Simon Willison, who built a plugin for it the same day, notes it was announced at OpenAI's DevDay the week before.

Most AI calls are decisions wearing a chat costume

A lot of production AI is not writing anything. It is deciding: is this spam, which queue does this ticket go to, how bad is this bug. Teams usually answer those with a chat model, a schema and some retry logic for when the model wanders off the format.

A dedicated endpoint changes two things. The answer comes back as a probability, so you set the threshold instead of parsing the model's mood. And because output is free, a long rubric costs you input only.

Yes, a classifier is not new. You could train one. Most teams do not have the labeled data or the time, which is why they reached for a chat model in the first place.

Before you move traffic

Check what the beta covers before you rebuild anything.

There is one model. If gpt-6-luna gets your categories wrong, there is nothing to swap to inside this API.

Images must be base64 inline, so a pipeline that passes image URLs needs a change.

The SDKs need recent versions: Python 3.26.0 or later and JavaScript 7.30.0 or later, among others listed in the guide.

On the same day, OpenAI also cut its API usage tiers from five to three, named Build, Launch and Grow. Organizations move up automatically as they buy credits. Check which tier you landed in before you load-test.

The practical test is small. Take a day of your classification traffic, send it through both paths, and compare accuracy, latency and cost on your own labels.

Watch the GA terms

The date to watch is general availability. Betas can change price and limits, and the two numbers that make this interesting are the free output and the 10x speed claim.

If both survive GA, routing through a chat model gets hard to justify.

Run the comparison now. Decide with your own numbers.