Decisions

POST /v1/decisions serves the OpenAI Decisions API. It evaluates shared input against up to 400 typed questions and returns an answer per question: the probability that a statement is true, one option from a fixed set, or a score on a rubric. It does not generate text, which makes it a fast, cheap fit for classification, routing and prioritization.

Requests and responses use OpenAI's wire format, so the OpenAI SDKs work by pointing base_url at LLM API. The model ID is gpt-6-luna-decisions: the decisions surface of GPT-6 Luna, listed separately from the gpt-6-luna chat model.

Quick start

curl -X POST "https://api.llmapi.ai/v1/decisions" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna-decisions",
    "input": "I was charged twice for my order. Please return the extra charge.",
    "questions": [
      {
        "type": "predicate",
        "name": "refund_requested",
        "instructions": "Does the customer ask for a refund?"
      }
    ]
  }'

{
  "model": "gpt-6-luna",
  "answers": [{ "type": "predicate", "name": "refund_requested", "probability": 0.95 }],
  "usage": {
    "input_tokens": 162,
    "input_tokens_details": { "cached_tokens": 0, "cache_write_tokens": 0 },
    "output_tokens": 0,
    "output_tokens_details": { "reasoning_tokens": 0 },
    "total_tokens": 162
  }
}

With the OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(base_url="https://api.llmapi.ai/v1", api_key=LLM_API_KEY)
decision = client.decisions.create(
    model="gpt-6-luna-decisions",
    input="Export fails in Safari but works in Chrome.",
    questions=[{
        "type": "score",
        "name": "severity",
        "instructions": "How severe is this issue?",
        "levels": [
            {"label": "Cosmetic"},
            {"label": "Workaround available"},
            {"label": "Fully blocked"},
        ],
    }],
)

Answers come back in question order and carry the question's name.

Question types

Type	Asks	Answer fields
`predicate`	How likely is this statement true?	`probability` (0–1)
`choice`	Which of 2–255 `choices` applies?	`choice`, `probabilities`, `confidence`
`score`	Where does the input sit on 2–10 ordered `levels`?	`score`, `probabilities`, `confidence`

Any question can come back as {"type": "refusal", "name": ...} while the others are still answered, so check each answer's type before reading it.

Input

input is either a string or an array of messages. User messages accept input_text, input_image (a public HTTPS URL or a base64 data URL) and input_json parts; developer and system messages are text only. Files, audio and file IDs are not supported.

"input": [{
  "role": "user",
  "content": [
    { "type": "input_text", "text": "Inspect the product in this photo." },
    { "type": "input_image", "image_url": "https://example.com/product.png" }
  ]
}]

Pricing and caching

Decisions are billed on input tokens only; output tokens are free. Repeated input of 1,024 tokens or more is cached automatically: cache reads are billed at the discounted cached-input rate and cache writes at the cache-write rate. prompt_cache_key and prompt_cache_options are forwarded to the provider unchanged. The request log shows the input, cached and cache-write tokens of every request; current rates are in the model catalog.

Providers

The request is served by OpenAI or Azure OpenAI; the response's model field names the model or Azure deployment that answered. When the first provider is rate limited or unavailable, the next one is tried. Pin a provider with a prefix, for example azure/gpt-6-luna-decisions; a pinned request is never moved to another provider. A provider that does not offer the Decisions API, such as AWS Bedrock, answers 400.

Notes

Ask several questions in one request: they share the same input, so one call is cheaper and faster than one per question.
Responses arrive in one piece. There is no streaming.
The gateway checks that model, input and 1–400 questions are present; everything else is validated by the provider, whose 400 message names the offending field.
/v1/systemone serves TypeSafe's System One models, which use a different request format.