Send OpenAI a support ticket and five queue names, and you get back one queue and a confidence of 0.93. That's the whole pitch of the Decisions API, which went to public beta on 6 October 2026. It bills $0.10 per million input tokens, nothing for output, and, as of this week, nothing is cached.
The short answer
Decisions is a POST /v1/decisions endpoint that runs only on gpt-6-luna and returns typed answers: a probability, a choice from your list, or a score against ordered levels. Input costs $0.10 per million tokens, output and cache operations cost nothing, and OpenAI says it's up to 10x faster than the Responses API. There's no caching, the probabilities look uneven in early tests, and general availability is promised "in the coming weeks".
What you send and what comes back
You pass a model, an input (text, or text plus images) and a list of questions. Each question has a type. A predicate returns the probability that a statement is true, a number between 0 and 1. A choice picks one value from options you define, with a confidence and a probability for each option. A score rates the input against ordered levels and returns a weighted average, which can land between two levels.
The docs' own example is a customer who writes "I was charged twice for my order". The answer comes back as billing, confidence 0.93, with a 0.95 probability on the billing option. No sentence is generated at all. That's the point, and it's why OpenAI can skip output pricing: there's almost no output to bill.
The limits are the part you'd want before building. Images have to be inline base64 data URLs, and hosted links are refused. Regional processing premiums and long-context multipliers apply, but the guide gives no percentages. It doesn't state how many questions fit in one request either. SDK floors are high: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 and Java 4.78.0. The beta supports zero data retention and HIPAA use for eligible customers, with data residency in the US and Europe.
What $0.10 in and nothing out really means
On the Responses API, Luna costs $0.10 in and $0.50 out, and a cached read is $0.01. Decisions keeps the input rate and drops the output charge. For a short prompt that's a small win, since a label is a handful of tokens anyway. The catch is the cache. Classification prompts are usually a long fixed rubric followed by a short variable item, which is exactly what caching was invented for.
Here's our arithmetic from list prices, not a measurement. Take a 3,000-token rubric and a 2,000-token ticket. On Decisions that's 5,000 tokens at $0.10 per million, so $0.0005. On Luna through the Responses API with the rubric cached, it's 3,000 tokens at $0.01, 2,000 at $0.10 and, say, 20 output tokens at $0.50, which is about $0.00024. Roughly half, and that ignores the one-off cache write, and that ignores cache write fees I couldn't confirm for Luna..125 per million on OpenAI’s price list, which adds about and that ignores cache write fees I couldn't confirm for Luna..0004 on the first call. So the faster endpoint is the dearer one whenever your prompt has a big stable prefix. It's cheaper only when prompts are short or never repeat.
The other comparison is Jev, the early access classifier we covered in Jev 1.13 bills $42 a billion tokens. $0.042 per million is less than half of Luna's $0.10 here. OpenAI's endpoint has the compliance paperwork and a general release path. Jev has the price, and, according to early users on OpenAI's forum thread, an edge on accuracy and speed for some tasks. That's one forum comment, not a benchmark, and OpenAI hasn't published accuracy numbers for Decisions.
What early testers found on the forum
Under the announcement, one developer tried a biased coin test and got heads 98% of the time against an expected 70%. They also reported that the order of the choices changes the probabilities. A single anecdote isn't proof, and I'd want to rerun it before repeating it as fact. But it matches a worry I'd have anyway: a number that looks like a calibrated probability is easy to trust and hard to check. If you route refunds or moderation on a 0.93, you need your own labelled set first.
My take: it's a good fit for cheap routing where a wrong answer costs a retry, and a poor fit for anything that treats the probability as a promise. I'd wait for GA pricing, since a beta rate that nobody committed to isn't a budget. The Luna row in our LLM API pricing tracker now notes the Decisions rate, and we'll update it when OpenAI moves.
Frequently asked questions
Which model does the Decisions API use?
Only gpt-6-luna for now. The docs list it as the single supported model during the public beta.
How much does the Decisions API cost?
$0.10 per million input tokens. There are no charges for output tokens, cache reads or cache writes. Regional processing premiums and long-context multipliers apply, and OpenAI hasn't published their values.
Does the Decisions API support caching?
A forum reply on the announcement says no caching is available. Cache operations aren't billed because they don't happen. For long, repeated rubrics, Luna on the Responses API with caching can cost less per request.
Can I send an image URL?
No. Images must be inline base64 data URLs. Hosted HTTP or HTTPS links and uploaded file IDs are refused.
Sources: OpenAI, Decisions API guide, API changelog and the public beta announcement of 6 October; independent report: AI Weekly. All read on 11 October 2026. The per-request costs are our arithmetic from list prices.
