Decisions

Answer typed questions about a piece of input. You send a state (the text or JSON to decide about) and a set of questions; each answer comes back as a label or a score with its probabilities and a confidence. The model answers each question with a single token, and only input tokens are billed.

POST https://inference.openrelay.inc/v1/decisions

The model that serves this family is openrelay/qwen3.6-35b-decide, built on Qwen3.6-35B-A3B by the Qwen team at Alibaba (Apache 2.0): a mixture-of-experts model with 35B parameters, 3B of them active per token. Check a model's supported_apis on GET /v1/models for decisions; a model that does not list it answers 400 unsupported_api. Decisions are online only: the Batch API does not accept /v1/decisions.

Question typesCopy Anchor Link

Every question has one of three types:

RequestCopy Anchor Link

Each question:

Any other field is rejected with 400 unknown_field, at the top level, inside a question, or inside noul criteria. That includes checkpoint, task, and lang. stream is dropped: a decision is always a single JSON response.

Use question ids made of ASCII letters, digits, _, and -. They come back as the keys of answers, and an error names them in its param field (see Errors).

How a question is answeredCopy Anchor Link

Each question is one prompt to the model: the state first, then the instructions, then the options lettered A, B, C, and so on, after a short fixed system prompt. The model produces one token, restricted to those letters, and the answer comes from the probabilities of the letters, renormalized over the options. A noul question is asked as A. yes / B. no, with the criteria.true and criteria.false descriptions attached to their sides when you send them, and noul is the probability of yes.

The whole state is read for every question. Nothing is truncated.

ResponseCopy Anchor Link

{
  "model": "openrelay/qwen3.6-35b-decide",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 1.0, "technical": 0.0 },
      "confidence": 0.9995
    },
    "urgency": {
      "type": "score",
      "score": 1.5995,
      "legend": { "0": "not urgent", "1": "soon", "2": "blocking" },
      "probabilities": { "0": 0.0314, "1": 0.3377, "2": 0.6309 },
      "confidence": 0.3028
    },
    "churn_risk": {
      "type": "noul",
      "noul": 0.9999,
      "confidence": 0.9999
    }
  },
  "usage": { "prompt_tokens": 539, "total_tokens": 539, "truncated_tokens": 0 },
  "openrelay": { "cost_nano_usd": 26950, "priced_at": "online" }
}

That is the response to the request in Examples.

LimitsCopy Anchor Link

BillingCopy Anchor Link

openrelay/qwen3.6-35b-decide costs $0.050 per 1M input tokens. It has no output rate.

The state is part of every question's prompt, so usage is the sum of each question's prompt tokens: the state, the instructions, the options, and a short fixed system prompt. A request with ten questions reads the state ten times, and bills for it. Asking several questions in one request saves round trips, not billed tokens.

usage.prompt_tokens in the response is the number you are billed on, and openrelay.cost_nano_usd what you are billed; they are the source of truth, not any estimate you compute from the request size. The response above bills 539 tokens, 539 x 50 = 26,950 nanoUSD. A request that fails with a non-2xx status is not billed. See Pricing & billing.

ErrorsCopy Anchor Link

A request the model rejects gets a typed error with a documented code and, in most cases, a param naming the field:

{
  "error": {
    "message": "too many options for one question",
    "type": "invalid_request_error",
    "code": "too_many_options",
    "param": "questions.department.criteria"
  }
}

The codes this family can return, all 400 with type invalid_request_error:

param is a dotted path into your own request, with [n] for array indices. It is present only when that path exists in what you sent, so a missing field (for example a question with no instructions) comes back without a param. It is also omitted when the path is longer than 128 bytes or passes through a question id that contains a dot, a space, or any character other than ASCII letters, digits, _, and -.

Every other error is the gateway's, listed on Errors: 401 for the key, 402 for the balance, 404 for an unknown model, 400 unsupported_api for a model that does not serve decisions, 413 for the body size, 429 when the model's queue is full, and 502 or 503 while the model is restarting or has no ready capacity. Retry 429, 502, and 503 with backoff (after the Retry-After header on a 429); do not retry a 400 unchanged.

What to expectCopy Anchor Link

Our own zero-shot measurement, one question per request, no examples in the prompt, on public labeled sets (2026-10-22):

What that means for you:

• Accuracy on your task is the number that matters. Measure it on your own labeled data before you rely on the answers.
• confidence tracks accuracy closely on English text and runs ahead of it in other languages, most of all in lower-resource ones. Pick any confidence threshold from your own labeled data, per language if your traffic mixes them.
• The model reads the whole state for every question, in any of the 201 languages Qwen lists. The limits and the price do not change with the language.

ExamplesCopy Anchor Link

curl https://inference.openrelay.inc/v1/decisions \
  -H "Authorization: Bearer $OPENRELAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrelay/qwen3.6-35b-decide",
    "state": {
      "subject": "Duplicate charge on invoice #4411",
      "body": "I was charged twice. If this is not fixed today I will cancel."
    },
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which department handles this?",
        "criteria": { "billing": "invoices, refunds", "technical": "bugs, outages" }
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent?",
        "criteria": ["not urgent", "soon", "blocking"]
      },
      "churn_risk": {
        "type": "noul",
        "instructions": "Does the user threaten to leave?"
      }
    }
  }'