Playbooks · Coding tasks
OpenAI Decisions API: Luna answers yes or no
Written November 8, 2026
OpenAI's Decisions API is generally available: Luna answers typed questions — predicate, choice, or score. What it fits, and what it doesn't.
What it is
The Decisions API is a dedicated endpoint, POST /v1/decisions, that evaluates text, images, and JSON and returns typed answers instead of prose. It runs only on gpt-6-luna, the only model currently supported. OpenAI announced it at DevDay in limited preview on September 29, 2026, opened a public beta on October 6, and made it generally available on October 22. The guide opens with that status: generally available, not a preview.
The point of the endpoint is the shape of the answer. A request has three parts — model, input as shared evidence, and questions — and the response is an answers array. Each question has a type, and the type decides what comes back:
- Predicate. Check a condition, such as whether a photo shows visible damage. The answer is a
probabilityfrom 0 to 1, the model's estimate that the condition is true. - Choice. Select one option from the values you supply, such as a department for a complaint. The answer is one of your supplied values, plus a
probabilitiesarray over the options and aconfidencefield. - Score. Rate an input against 2 to 10 ordered levels, such as issue severity. The answer is a probability-weighted average of the level indices, so it can fall between levels, plus the per-level probabilities and a confidence field.
A request can carry up to 400 questions, up from 200 during the beta, and independent questions can share one input. Questions can depend on earlier answers only by sending separate requests — check for damage first, then ask for a repair category. The guide's named uses are classifying content, routing requests, and choosing an agent's next action. OpenAI says the endpoint returns typed answers about 10 times faster than the Responses API with the same model.
What it costs
With gpt-6-luna, uncached input costs $0.10 per 1 million tokens, and output tokens are not charged. Prompt caching reuses a shared input prefix of at least 1,024 tokens: cache hits receive discounted input pricing and cache writes are charged. Regional processing premiums and long-context input pricing multipliers apply. The guide notes the API supports Zero Data Retention and HIPAA use for eligible customers, with data residency in all supported regions and regional processing in the United States and Europe.
What it is not
- No chat, no prose. Decisions returns typed answers; it writes no prose, no explanation, and no extracted fields. When the job needs a generated object that follows your own JSON schema, or a model requesting a tool call, the guide points back to Structured Outputs with the Responses API or to function calling.
- Not a decision-maker. The guide is explicit: Decisions returns typed answers, and your application runs any actions those answers select. The endpoint picks the answer; the action, the threshold, and the consequences stay with your code.
- Not a control by itself. The answer set is enforced by the response format — a choice answer can only be one of the values you supplied — but the quality of the answer still depends on how the question is written. The guide's advice is to write questions around observable criteria, give choices distinct meanings, and define score levels so adjacent levels have distinct criteria. A vague question gets a confident-sounding probability, not a good decision.
- Not open-ended. Files, audio, and item references are not supported as inputs. A question that needs them does not fit this endpoint.
What to check before using it
- Try one recurring decision in the Playground first. OpenAI provides a Decisions Playground for experimenting with questions and inputs before writing code. Pick a decision your application already makes often, define the question and the possible answers, and test it on examples from real usage.
- Set thresholds from labelled examples. Predicate answers are probabilities, and choice and score answers carry a confidence field. The guide's advice is to choose thresholds based on the cost of false positives and false negatives, using labelled examples from your application — not to take 0.5 as a default.
- Include a fallback option. The guide suggests a choice such as
"other"for inputs the categories do not cover, so the result can go to a general review queue instead of forcing a wrong category. - Do not rebuild it with prompts. The temptation is to ask a chat model the same question and parse the answer. That is the failure the endpoint exists to remove: a parsed answer can invent a category, while a choice answer cannot leave the set you supplied.
Limits
- This page reads OpenAI's Decisions guide. It reports no API call here — no key, no request, no measured latency — so the speed figure is OpenAI's, not a measurement, and this page says nothing firsthand about accuracy on any real workload.
- The availability history on this page comes from the guide and OpenAI's release record: limited preview at DevDay, public beta October 6, generally available October 22. The guide is the current record; earlier coverage describing the API as a limited preview is out of date.
- Only
gpt-6-lunais supported. A decision that needs a different model, or a question type beyond predicate, choice, and score, does not fit this endpoint.
Related playbooks
- OpenAI's Agents API: the harness is now hosted
- Claude API effort control: change gear mid-stream
- Gemini API now takes Office files and returns files
- Gemini Decision: a model and API for choices
Sources
Back to: Coding playbooks hub · Also: OpenAI's Agents API: the harness is now hosted · Claude's watermark now covers older models — mostly