JSON mode answers one question: will json.loads() succeed? It does not answer the second: is this the object my code expects?
How it works
In the Chat Completions API you turn it on with one field:
{"response_format": {"type": "json_object"},
"messages": [
{"role": "system", "content": "Return JSON with keys name and email."},
{"role": "user", "content": "Reach me at Sara, sara@example.com"}]}
OpenAI’s API reference calls this “the older JSON mode, which ensures the message the model generates is valid JSON.” Two rules come with it. You must ask for JSON in a message: OpenAI’s guide says the API “will throw an error if the string ‘JSON’ does not appear somewhere in the context.” And the shape is up to the model. The guide states that JSON mode “will not guarantee the output matches any specific schema.”
Structured output is the stricter option. It passes a JSON Schema with {"type": "json_schema", ...}, and the model is constrained to it.
JSON mode (json_object) |
Structured outputs (json_schema) |
|
|---|---|---|
| Valid JSON | yes | yes |
| Required keys and types | no | yes |
| You supply a schema | no | yes |
| Must mention “JSON” in prompt | yes (OpenAI) | no |
OpenAI recommends “always using Structured Outputs instead of JSON mode when possible.”
Why it matters
- Validation is still yours. With JSON mode, a missing key or a number sent as a string parses fine and fails later. Validate against a schema in code.
- Truncation. If the reply hits
max_tokens, the JSON is incomplete. Reasoning models are more exposed, since thinking uses the same budget. - Fallback. OpenAI’s guide lists JSON mode on older models where
json_schemais not available. Support varies by provider and model, so JSON mode plus your own validation is the fallback when a schema is not accepted.
On BroadNet Inference
BroadNet’s API reference lists structured output for GLM 5.3 and Qwen3.8 27B. It does not say whether json_object or json_schema is the supported response_format, or whether DeepSeek V4.1 Flash and BroadNet Gemma 4 VL 26B accept either. Test both on your model. To write the schema, use the JSON Schema generator. The post JSON mode vs json_schema covers repair and monitoring, and the invoice guide shows JSON extraction from images.