POST /v1/decide

Ask typed questions about some text (and optionally images) and get a calibrated answer for each: a choice with a probability for every option, a score on a scale, or a yes/no probability. For a guided tour, see Decision mode.

Requests to the Flux decision model are processed by Cloudflare Workers AI, which acts as Flux's sub-processor and handles them under Cloudflare's own terms. Flux keeps no copy of your request or the answer; each call is billed from its input token count only.

Only paid plans can use it. A key on a free plan, or a key limited to other models, is refused with 403 before anything is processed.

Request#

Other top-level fields and other question fields are ignored.

Every question has a type and non-empty instructions, then the part that depends on the type:

curl https://api.fluxrouter.ai/v1/decide \
  -H "Authorization: Bearer $FLUX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "flux-decide",
    "state": "Checkout returns a 500 error since this morning.",
    "questions": {
      "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
          "billing": "Payments, invoices and refunds",
          "technical": "Errors, bugs and outages",
          "sales": "Pricing and new plans"
        }
      },
      "urgent": { "type": "noul", "instructions": "Is this urgent?" },
      "severity": {
        "type": "score",
        "instructions": "How severe is the problem?",
        "criteria": ["minor", "moderate", "major", "critical"]
      }
    }
  }'

Response#

200 with one answer for every question you asked, under the same id:

{
  "model": "flux-decide",
  "answers": {
    "team": {
      "type": "choice",
      "choice": "technical",
      "probabilities": { "billing": 0.0197, "technical": 0.9728, "sales": 0.0075 },
      "confidence": 0.9202
    },
    "urgent": { "type": "noul", "noul": 0.9279 },
    "severity": {
      "type": "score",
      "score": 2.4414,
      "probabilities": { "0": 0.0122, "1": 0.0284, "2": 0.4653, "3": 0.4941 },
      "legend": { "0": "minor", "1": "moderate", "2": "major", "3": "critical" },
      "confidence": 0.2822
    }
  },
  "usage": { "input_tokens": 356 }
}

Threshold on the probability, never on confidence#

To decide whether to trust an answer, compare probabilities[choice] with your threshold (for a noul question, compare noul). Do not use confidence. It is not the probability of the chosen option: in the example above the chosen option has a probability of 0.9728 and a confidence of 0.9202, and the score question has a most likely level of 0.4941 against a confidence of 0.2822.

Response headers#

Limits#

The state limit#

The model reads the first 2,048 tokens of the state and silently drops the rest. A long text would therefore be answered from its beginning alone, and the answer could be wrong without any sign of it. So Flux refuses instead: a state over the limit is answered with 422 state_too_long and nothing is billed. Shorten the state, or send only the part the questions are about. Images do not count toward the state limit.

Pricing#

Each call is billed once, from its input tokens only, at 0.1008 USD per million input tokens. A typical decision of about 450 input tokens costs about 0.00005 USD. The same price applies on every paid plan. A call that fails is not billed.

Errors#

The error body is the same on every endpoint that returns the standard shape:

{ "error": { "message": "The state is longer than 2048 tokens", "type": "invalid_request_error", "code": "state_too_long", "param": "state" } }

The message is fixed text written by Flux. It never repeats your state, instructions, option text, question ids or images. param, when present, names the request field at fault: model, state, questions, images or body.

A 403 means this key or plan cannot use decision mode; a 402 means a money check stopped the call. Every code is also in the error reference.