JSON Mode: Definition and Examples

JSON Mode is a feature offered by language model providers (such as OpenAI, Anthropic, or Google) that constrains the model's output to syntactically correct JSON format. When activated, the model ensures that its response constitutes a valid JSON object, eliminating common parsing issues related to natural language responses.

In practice, JSON Mode solves a fundamental problem of using LLMs in production: output format reliability. Without this constraint, a model may add explanatory text before or after the JSON, use incorrect quotation marks, or produce malformed JSON. JSON Mode guarantees that each response can be directly parsed by a standard JSON interpreter without additional processing.

It is important to distinguish JSON Mode from Structured Output (or JSON Schema). JSON Mode only guarantees that the output is valid JSON, but does not control the exact structure of the data (which keys, which value types). Structured Outputs go further by allowing you to specify a precise JSON schema that the model must follow. JSON Mode is therefore a format constraint, while Structured Outputs are a structure constraint.

For best results with JSON Mode, it is recommended to explicitly describe the expected JSON structure in the prompt, by providing an example or a schema. The model will then respect both the technical constraint (valid JSON) and the semantic constraint (desired structure) defined in your instructions.