JSON Mode

Interface tools published

Also known as: JSON Output Mode, JSON Response Format

Definition

An API setting that constrains model output to be valid JSON, without specifying a particular schema. JSON mode guarantees parseable JSON but not any specific structure—the model chooses the keys and values. It's a weaker guarantee than structured output (which enforces a schema) but simpler to use.

What this is NOT

Not structured output (JSON mode doesn't enforce a schema)
Not guaranteed to have specific fields (only guarantees valid JSON syntax)
Not function calling (that's for tool invocation, not output formatting)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

An API parameter (e.g., `response_format: { type: "json_object" }`) that ensures the model outputs valid JSON. Introduced before full structured output support as a middle ground between free text and schema constraints.

Sources: OpenAI JSON mode documentation, Anthropic JSON output

Examples

response_format: { type: 'json_object' } in OpenAI API
Model outputs {"answer": "Paris", "confidence": 0.95} without schema
Extracting key-value pairs with flexible structure

Counterexamples

Things that might seem like JSON Mode but are not:

Structured output with a defined schema (that's stricter)
Free-form text that happens to contain JSON
Function calling (different mechanism)

Relations

specializes structured-output (JSON mode is a weak form of structured output)
overlapsWith function-calling (Both produce structured data, different purposes)

Implementations

Tools and frameworks that implement this concept:

GPT-4 Turbo secondary