OpenAI's Decisions API Charges $0.10 per Million Input Tokens and Nothing for Output

OpenAI's Decisions API bills $0.10 per million uncached input tokens and charges nothing for output, according to the endpoint's documentation, which now lists the product as generally available on a single model, gpt-6-luna.

That is a change from October 6, when OpenAI opened the endpoint as a public beta and said general availability would follow "in the coming weeks," as Let's Data Science reported. The docs raise the cap from 200 questions per request in the beta to 400.

What the endpoint returns

The call is POST /v1/decisions. A developer sends text, images or JSON plus a list of questions, and gets back typed answers instead of a paragraph. OpenAI's documentation defines three types.

Type	What you ask	What comes back
predicate	Is this condition true?	A probability from 0 to 1
choice	Which of these fixed options?	The chosen value, plus probabilities and confidence
score	How does this rate on 2 to 10 levels?	A probability-weighted score, plus probabilities and confidence

A response can also be a refusal, which the docs show in their examples. The practical task is routing: decide whether a support ticket is billing or abuse, or whether an uploaded image breaks a rule, and let ordinary code act on the number.

The price in practice

Uncached input costs $0.10 per million tokens. Cache hits are discounted, cache writes are charged, and regional processing premiums and long-context multipliers apply, per the docs. Output tokens are free.

The arithmetic: 10,000 requests of 1,000 input tokens each is 10 million tokens, or $1.00 uncached. A reusable cached prefix must be at least 1,024 tokens long to qualify.

1. Public beta200 questions
2. Current docs400 questions

Source: OpenAI Decisions API documentation, accessed 2026-10-23

What it cannot take

The input list is narrow. Per the docs, files, audio, file_id inputs and item references are not supported. Let's Data Science, summarising the beta launch material, adds that function calls, function-call outputs and non-user roles are also excluded, with a maximum of 128 image parts per request. The docs do not repeat those last limits, so check them against the live page before building on them.

The docs do not state a rate limit. For regulated users they list Zero Data Retention and HIPAA availability for eligible customers, regional processing in the United States and the EEA plus Switzerland, and Private Link.

The speed claim has no workload behind it

OpenAI says the endpoint is about 10 times faster than the Responses API. Let's Data Science says the launch materials give no public workload definition or latency distribution, and recommends measuring end to end on your own data. The Terminal found no independent benchmark of the endpoint. Treat 10x as a vendor-supplied figure.

The probabilities deserve the same caution. The docs describe confidence and probabilities as outputs, but the pages we read do not say how well calibrated they are, or what threshold suits which task. A choice answer reporting 0.92 is the model's own estimate, not a measured accuracy.

For comparison with OpenAI's other platform moves, see our report on GPT-6 reaching Google Cloud in private preview with no pricing and the OpenAI CLI's mismatched README and docs. Usage reporting sits at GET /v1/organization/usage/decisions and needs an admin key with the api.usage.read scope.

The open date is general availability itself: OpenAI's beta announcement promised it in weeks, per Let's Data Science, and the docs already read as if it has happened, with no dated announcement we could find.