Structured Outputs & JSON Mode (2026): Reliable JSON From LLMs

Why Structured Outputs

If your app parses LLM output, free-form text is a liability — a stray sentence or markdown fence breaks JSON.parse. Structured outputs make the model return valid, schema-shaped JSON every time, so extraction, classification and pipelines are reliable.

Three Levels of Reliability

1. Prompt for JSON — ask for JSON and give an example. Easy, but not guaranteed.
2. JSON mode — the model is constrained to emit syntactically valid JSON.
3. Schema-constrained — the output must match a JSON schema you supply (strongest guarantee).

Use the strongest level your chosen model supports — confirm options in the docs.

Example

payload = {
  "model": "gemini-3.5-flash",
  "messages": [
    {"role": "system", "content": "Extract fields as JSON only. No prose, no markdown."},
    {"role": "user", "content": "Invoice #A-190 from Acme, total $4,200, due 2026-08-01."}
  ],
  "response_format": {"type": "json_object"}   # enable JSON mode (name varies by model)
}
# expected: {"invoice_no":"A-190","vendor":"Acme","total":4200,"due":"2026-08-01"}

Then validate before trusting it:

import json, jsonschema
schema = {
  "type": "object",
  "properties": {
    "invoice_no": {"type": "string"},
    "vendor": {"type": "string"},
    "total": {"type": "number"},
    "due": {"type": "string"},
  },
  "required": ["invoice_no", "total"],
}
data = json.loads(resp_text)
jsonschema.validate(data, schema)   # raises if the shape is wrong

Make It Bulletproof

• Always validate against a schema; don't assume.
• Retry on invalid — reissue with "Your previous output was invalid JSON; return only valid JSON matching the schema." One retry fixes most cases.
• Say "JSON only, no markdown" in the system prompt to avoid ```json fences.
• Keep schemas flat where you can; deeply nested schemas are harder for models.
• Lower temperature for extraction tasks.

Great For

Data extraction, classification/tagging, function arguments (pairs with function calling), and any step feeding downstream code. For cheap high-volume extraction, DeepSeek V4 Flash or Gemini Flash are strong; see cheapest LLM API.

Bottom Line

Enable JSON/schema mode, tell the model "JSON only," validate against a schema, and retry once on failure. Do that and LLM output becomes a dependable data source. Start from the Python quickstart.

Start free with a $1 credit and lock down your JSON.