Structured outputs

Force JSON output with response_format. Two modes: json_object (any valid JSON) and json_schema (strict schema conformance). Examples in curl and Python.

The response_format parameter forces the model to return valid JSON. Two modes are supported.

json_object mode

The model guarantees the output is valid JSON. You define the structure in the system prompt.

curl https://getvibeapi.com/v1/chat/completions \
  -H "Authorization: Bearer $GETVIBE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6.1-sol",
    "response_format": { "type": "json_object" },
    "messages": [
      {"role": "system", "content": "Reply with JSON with fields: name (string), score (number)."},
      {"role": "user", "content": "Rate this answer: very good"}
    ]
  }'

json_schema mode

Strict JSON Schema compliance — the model fills exactly the specified fields. Supported by GPT-5.x and compatible models.

curl https://getvibeapi.com/v1/chat/completions \
  -H "Authorization: Bearer $GETVIBE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6.1-sol",
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "review",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "name": { "type": "string" },
            "score": { "type": "number", "minimum": 1, "maximum": 10 },
            "summary": { "type": "string" }
          },
          "required": ["name", "score", "summary"],
          "additionalProperties": false
        }
      }
    },
    "messages": [
      {"role": "user", "content": "Rate this product: fast delivery, careful packaging"}
    ]
  }'

With strict: true the gateway routes the request only to model variants that actually honour the schema, and it does so on every attempt — including a switch to a backup supplier. If no suitable variant is available right now you get an explicit structured_output_unsupported error rather than plain text where JSON was requested. That guard is built from measurements and covers only the variants we have measured: on an unmeasured variant the supplier may still discard the schema silently.

Models differ on how well they hold JSON — check yours. Measured 2026-09-12 on deepseek-v4-flash-0731, 10 calls per cell: json_object parsed 10/10 on the :cxa and :cxbc variants and 8/10 on :network, while a strict json_schema parsed 0/10 on both variants (:cxa answers 400, :cxbc answers 200 with plain text) and 3/10 on :network. DeepSeek's own API has no json_schema mode — use json_object, and when you need exactly one shape, a forced function call (tools + tool_choice) returned parseable arguments 10/10 on every variant. Per-variant numbers live in the structuredOutput field of GET /v1/model-capabilities. Each is a one-off reading stamped with its date (measurement: "point-in-time"), not a promise: the supplier behind a variant can change, and we do not re-measure on a schedule.

Example response

{
  "id": "chatcmpl-...",
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "{\"name\":\"product X\",\"score\":9,\"summary\":\"Fast and convenient\"}"
    },
    "finish_reason": "stop"
  }],
  "usage": { "prompt_tokens": 42, "completion_tokens": 18, "total_tokens": 60 }
}

← PreviousMessages & Responses Next →Function calling