Decisions (Beta)
Decisions
Ask typed questions about text or images and get calibrated probabilities back, in the OpenAI Decisions format
POST
cURL
Until Monday, October 26, 2026, this endpoint took TypeSafe’s System One format. That format is served at
/v1/systemone, so a client written against it keeps working by changing its path.Overview
Unlike the Responses or Chat Completions APIs, a decisions model doesn’t generate text. You send it an input and a list of typed questions, and it returns a calibrated probability for each one. Use this when your code needs a number it can branch on, such as “how likely is this support ticket to be a bug report” or “which of these three categories does this transaction belong to,” rather than a paragraph it has to parse. In/v1/models, a decisions model is one whose provider entry has supports.endpoint.decisions: true. OpenAI’s gpt-6-luna also answers the chat endpoints. TypeSafe’s jev-1.13.0 and jev-1.14.0, Cloudflare’s clef and clef-flash, Inception’s mercury-decide, Baseten’s elephant-1 and mammoth-1, and Google’s gemini-decision-1 answer decisions only. All of them take the same request, whichever format their provider speaks natively.
Input
input is the content the questions are judged against, either a string or an array of user messages. A message’s content is a string or an array of input_text and input_image parts. Images are data URLs. An image in a format the model doesn’t read (see Models) is a 400, never silently dropped. Images are billed as the provider counts their tokens. Other roles, http image URLs and files are also a 400.
Question types
Each question has atype, a required instructions string (empty is allowed) and an optional name. TypeSafe’s box question is served on /v1/systemone only. Names must be unique within the request. Answers come back in question order, each carrying the name you sent (or null).
predicate
A yes/no judgment. Returns a probability between 0 and 1 that the answer is yes.
choice
Selects one of 2 to 255 choices. Each has a value (a string or a boolean, unique within the question) and an optional description. Returns the chosen value, a probability for every choice in the order you sent them, and a confidence.
score
A position on 2 to 10 ordered levels, lowest first. Each has a label and an optional description. Returns a probability-weighted score, a probability for every level with its 0-based value and label, and a confidence.
Answers
The response is{ "id", "model", "answers", "usage", "cost" }. answers has one entry per question, in question order. Each entry’s type matches the question that produced it.
- A
predicatereturns{ "type": "predicate", "name": "is_bug_report", "probability": 0.87 }. There’s noconfidencefield, because the probability is the answer. - A
choicereturns{ "type": "choice", "name": "category", "choice": "shipping", "probabilities": [{ "value": "shipping", "probability": 0.91 }, { "value": "technical", "probability": 0.06 }, { "value": "account", "probability": 0.03 }], "confidence": 0.91 }. - A
scorereturns{ "type": "score", "name": "urgency", "score": 1.7, "probabilities": [{ "value": 0, "label": "Low", "probability": 0.05 }, { "value": 1, "label": "Medium", "probability": 0.2 }, { "value": 2, "label": "High", "probability": 0.75 }], "confidence": 0.78 }. - A
refusal,{ "type": "refusal", "name": "category" }, means the model couldn’t answer from the input. We’ve seen it when a question asks about an image that isn’t attached. It carries no probability.
choice and score answers, confidence summarizes how concentrated the probability distribution is. It isn’t a statement that the chosen answer is correct. Several acceptable options spread probability across them and lower confidence even when the model is judging soundly. The order of the options you send can change the answer, on every model, so keep it stable between calls you compare.
id is the provider’s request id, the Response ID the logs page shows. usage is in the Responses shape. Decisions are priced on input tokens alone, and output tokens aren’t charged. cost is in dollars, in the same shape the chat endpoints return. Fields a provider adds to an answer beyond the documented schema reach you rather than being stripped, so allow for unknown fields on individual answers.Models
gpt-6-luna, jev-1.13.0, jev-1.14.0, clef, clef-flash, mercury-decide, elephant-1, mammoth-1 and gemini-decision-1, with their aliases and provider/ prefixes (openai/gpt-6-luna, typesafe/jev-1.13.0, deepinfra/clef). A prefix that isn’t a provider of the model, such as openai/jev-1.13.0, is a 400. List them with GET /v1/models?supports.endpoint.decisions=true.
Pricing is each provider entry’s input rate, listed in
/v1/models. A request over gpt-6-luna’s limit gets a 400 before it is sent. One over Jev’s comes back as TypeSafe’s max_tokens_exceeded 400, and an image outside Jev’s size limits as TypeSafe’s 400. One over Clef’s, Mercury Decide’s, Elephant’s, Mammoth’s or Gemini Decision 1’s comes back as a 400 with code context_length_exceeded, as on the chat endpoints. Google reads a gemini-decision-1 request’s questions in the alphabetical order of their names rather than in array order, and that order can shift answers slightly. The TypeSafe-Danger-Zone-Long-Context header is forwarded to Jev and ignored by gpt-6-luna. Decisions requests share your key’s token rate-limit windows with every other endpoint.
Aliases such as
jev-latest move when the provider ships a new release, so the answers behind them can change without a change on your side. The response’s model field reports the versioned ID that answered. If you’ve tuned confidence thresholds against a version, pin its ID and move on your own schedule.Redaction, ZDR and HIPAA
Redaction, on the key or the request, isn’t supported here: the request gets a400 before it reaches the provider. Use a key without redaction enabled for Decisions traffic.
ZDR and HIPAA mode apply as on the chat endpoints. A provider that can’t meet the policy gets a 422. TypeSafe offers neither zero data retention nor a Business Associate Agreement, and neither does Google AI Studio, which serves gemini-decision-1.
A request can carry the same settings as on the chat endpoints, including routing, except redact. See Settings.
BYOK
A stored key for the model’s provider works here as a BYOK key does on the chat endpoints. The request runs on your key,cost.byok is true and no credits are charged. See BYOK.
Errors
See Error Handling for the general error response format.
Related Documentation
System One
The same models in TypeSafe’s format
Writing questions
TypeSafe’s guide to state, instructions and criteria
Create Response
Generate text with the Responses API
Error Handling
Understand error responses across the API
Body
application/json
Response
Default Response
Last modified on November 12, 2026

