OpenAI Decisions API: what it is and what it costs

The OpenAI Decisions API is a new endpoint that answers questions you define in advance with probabilities, not prose: is this true, which of these options fits, how bad is it on a scale you set. It runs only on GPT-6 Luna, costs $0.10 per million input tokens, and charges nothing for output.

OpenAI first showed it at DevDay. On September 29, @OpenAIDevs said it would let developers "classify content, route requests, or choose an agent's next action," but only in limited preview. A week later, on October 6, the same account announced the public beta, open to every developer, and claimed it "makes decisions up to 10x faster than GPT-6 Luna through the Responses API." That post has passed 1.8 million views.

What the Decisions API does

According to OpenAI's guide, you send a request to /v1/decisions with some evidence (text, images, or both) and a list of named questions. Every question has one of three types:

• A predicate question returns the probability, from 0 to 1, that a statement is true ("Does this photo show visible damage?").
• A choice question picks one option from a list you supply and returns a probability for each option plus a confidence number ("Billing, scheduling, or complaint?").
• A score question rates the input against ordered levels you define. The result is a probability-weighted average, so it can land between two levels.

The answer has to be one of your options, so you never get a chatty paragraph that breaks the code reading it. The sample uses in the guide are ordinary: flag damaged product photos, send complaints to the right department, rank how severe an issue is, filter content by relevance.

Some limits matter right now. Images have to be sent inline as base64, and hosted image links or uploaded files get refused. An answer can come back as a refusal, which your code has to handle. The docs give no maximum number of questions per request, and when one decision depends on another, you make two separate calls.

What it costs

Input is $0.10 per million tokens. Output, cache reads and cache writes are free. The fine print in the docs says regional processing premiums and long-context multipliers still apply, and OpenAI hasn't promised the beta price will hold at general availability, which it expects "in the coming weeks."

For a sense of scale, 10,000 customer emails of about 500 tokens each comes to 5 million tokens, or roughly 50 cents. The docs also say the endpoint supports Zero Data Retention and HIPAA for eligible customers, with data residency in the US and Europe.

How it compares to Jev and Perplexity

OpenAI is late to this category. TypeSafe launched Jev, a model that returns odds and never writes, in mid-September. Perplexity followed on October 2 with its own Decisions API at $0.04 per million input tokens, built on an open model you can host yourself. OpenAI's price is about two and a half times Perplexity's. What OpenAI has going for it is the platform most developers already use, plus the compliance paperwork bigger customers need.

What early testers found

The first reports from developers are mixed, and they're small samples, so don't treat them as benchmarks. On OpenAI's developer forum, a user named jefff ran a few hundred test questions through both Luna and Jev. They found Luna produced about three times as many confident wrong answers on nuanced calls, took about 1.6 seconds per turn against Jev's 0.3, and cost two to three times as much.

Another user, platypus, tried a biased coin that lands heads 70% of the time. Asked as a predicate, the API said heads 70% of the time, which is right. Asked as a choice, it picked heads 98% of the time. In other words, how you word the question changes the answer, and the choice format leans toward the favorite.

I think that second finding is the most useful thing to know before building on this. Use predicates when you care about the actual odds, and test your thresholds on real examples from your own business, not the demo.

What this means for a small business

A lot of the AI work a small business could use is sorting, not writing: is this lead real, is this message urgent, which tech gets this job, does this invoice look off. Until recently you'd do that with a chat model and careful prompting, which was slow and hard to predict. Now three companies sell tools built just for the sorting, at prices close to zero, and a form that tags and routes every inquiry as it arrives costs pennies a month to run.

Choosing between them is the quick part. Most of the effort goes into writing down what each decision should be and what a wrong answer looks like. If you want help finding the sorting jobs in your own week that are worth handing off, New Face Design's free process audit starts there.