Decision models

A decision model answers typed questions — the probability that a statement is true, one option of a closed set, a level on a rubric — and generates no text. TypeSafe's Jev was the first; OpenAI, Cloudflare, Perplexity, Inception, Liquid AI, Upstage and Microsoft now ship one. INFRO runs every decision feature on whichever of them you connect, translating each provider's wire into one shape, so a policy, a guardrail check or a route classifier names a model and nothing else changes.

The models

An OpenRouter connection reaches every one of them too, on OpenRouter's own System One endpoint. Connect both a provider and OpenRouter, or two providers' models in a route, and a decision fails over between them like any other request.

Name a model by its slug in a decision policy, a guardrail's classifier checks, a route's prompt classifier, POST /v1/rerank, POST /v1/systemone or POST /v1/decisions. On a decision endpoint, gpt-6-luna and openai/gpt-6-luna mean GPT-6 Luna's decisions; on chat completions they mean the chat model. A decision model named on chat completions is refused with 400 decision_model.

POST /v1/decisions in OpenAI's dialect

OpenAI's SDK (client.decisions.create) works against INFRO with the base URL and the key changed. The request runs on your own connections with every INFRO policy applied — key scopes, budgets, routes and failover, the request log — and the model need not be OpenAI's: any decision model answers in OpenAI's shape.

import os
from openai import OpenAI

# The INFRO gateway key, not the OpenAI key: that stays in your INFRO connection.
client = OpenAI(base_url="https://api.infro.io/v1", api_key=os.environ["INFRO_API_KEY"])

decision = client.decisions.create(
    model="gpt-6-luna",  # or any decision model: "cloudflare/clef-flash", "typesafe/jev-1.13"
    input="I was charged twice for my order.",
    questions=[
        {
            "type": "choice",
            "name": "department",
            "instructions": "Which department should handle this complaint?",
            "choices": [
                {"value": "billing", "description": "Payments, invoices, and refunds."},
                {"value": "technical", "description": "Problems using the product."},
                {"value": "other", "description": "Requests outside these categories."},
            ],
        }
    ],
)
print(decision.answers[0])

{
  "id": "req_…",
  "object": "decision",
  "model": "gpt-6-luna",
  "answers": [
    {
      "type": "choice",
      "name": "department",
      "choice": "billing",
      "probabilities": [
        { "value": "billing", "probability": 0.95 },
        { "value": "technical", "probability": 0.03 },
        { "value": "other", "probability": 0.02 }
      ],
      "confidence": 0.93
    }
  ],
  "usage": {
    "input_tokens": 61,
    "input_tokens_details": { "cached_tokens": 0, "cache_write_tokens": 0 },
    "output_tokens": 0,
    "output_tokens_details": { "reasoning_tokens": 0 },
    "total_tokens": 61,
    "cost": 0.0000061,
    "price_source": "list"
  },
  "route": "primary",
  "route_name": null,
  "provider": "openai",
  "connection": "conn_…"
}

• input is a string, or user messages of input_text and input_image parts (at most 128 images). An image reaches only a model that sees images; a request whose model reads text only moves to its next fallback, and with none left is refused with 400 images_unsupported. Every decision model takes images only as inline data URLs: Cloudflare, Perplexity and Liquid AI are refused a link with 400 image_url_unsupported before any call, and OpenAI's Decisions API refuses one with its own 400.
• questions are predicate, choice and score, as OpenAI documents them. On a System One model a predicate is a yes/no question, a choice's options are keyed by their text (a boolean value comes back as the boolean), and a score's level reads label: description.
• A question no model answered readably comes back as OpenAI's {"type": "refusal"}; an unnamed question is answered with name: null.
• safety_identifier is passed to OpenAI and to no other provider.
• Responses add id, usage.cost, usage.price_source, route, route_name, provider and connection to OpenAI's fields.
• POST /v1/decisions/{policy} is a saved decision policy: a different route under the same path.

Images on POST /v1/systemone

POST /v1/systemone takes TypeSafe's body and, for the models that see images, a top-level images array — data URLs, or {content_type, base64} — the extension Cloudflare and Liquid AI add to it. Each provider receives the images in its own form.

Errors