How is this different from the old JSON mode?

The older json_object mode only guaranteed valid JSON, not a valid shape — the model could return any keys. Structured outputs with json_schema and strict true guarantee the response matches your exact schema, so missing or extra fields become impossible rather than something you defend against.

What are the rules for strict mode?

In strict mode every property must be listed in the required array and additionalProperties must be set to false. Optional fields are not allowed as such — to make a field optional you instead allow null as one of its types. The builder above enforces the required and additionalProperties rules for you.

Which models support structured outputs?

GPT-4o from the 2024-08-06 snapshot onward and later models support json_schema with strict mode. Older models only support the looser json_object mode. Always pin a dated model string so a future change does not silently drop the guarantee.

Can I use nested objects and arrays?

Yes. Schemas can nest objects and arrays to a reasonable depth, and every nested object must also follow the strict-mode rules. The builder shows flat fields including string arrays to keep things clear, but the same response_format structure scales to nested schemas.

Does it cost more or slow things down?

There is no surcharge for structured outputs and latency is essentially unchanged. You do pay for the tokens of the schema you send and the JSON the model returns, but in practice it saves money by eliminating the retry loops that unstructured JSON parsing usually needs.

How to Use Structured Outputs with OpenAI

The problem structured outputs solves

Asking a language model to “return JSON” and then parsing the result is one of the most fragile patterns in AI engineering. The model adds a markdown fence, omits a field, invents an extra key, or trails a comment — and your parser throws. Teams paper over this with retry loops, regex stripping, and try/catch blocks. OpenAI’s structured outputs removes the whole class of bugs: you describe the exact shape you want with a JSON Schema, and the model is constrained to produce output that matches it. Build that request body with the tool above.

How it works

You pass a response_format of type json_schema containing your schema and strict: true. Under the hood, OpenAI compiles your schema into a grammar that constrains decoding, so the model literally cannot emit a token sequence that violates the structure. The returned message.content is then a JSON string you can parse directly with full confidence.

Strict mode has two non-negotiable rules that trip people up:

Every property must be in required. There is no implicit optional. To express “this field may be absent,” allow null as one of its types and mark it required anyway.
additionalProperties must be false. This is what stops the model from inventing keys you didn’t ask for.

The builder enforces both rules automatically, so the body it produces is valid on the first try.

Tips and patterns

Pin a dated model. Use gpt-4o-2024-08-06 or newer; older snapshots only support the weaker json_object mode that doesn’t guarantee the shape.
Use it for extraction and tool-like APIs. Structured outputs shine when you pull fields out of unstructured text, classify into a fixed enum, or return data an application will consume programmatically.
Prefer enums over free strings for categories. Adding an enum to a string property guarantees the model picks one of your allowed values, which is far more reliable than instructing it in the prompt.
Keep schemas reasonable. Very deep or huge schemas cost tokens and can hit limits; model the data you actually need, not everything you might.