← Voltar aos guias

LLM Structured Output and JSON Mode in 2026

LLMs are excellent at producing useful text, but free-form text is difficult for software to consume safely. A model may add commentary, omit a field, return invalid JSON, or change a value's type between requests. Structured output and JSON mode address this problem by constraining the response to a machine-readable format.

In 2026, the reliable pattern is not simply “ask for JSON.” Define a contract, request constrained output, validate it in your application, and recover predictably when the response still fails.

JSON mode versus structured output

JSON mode tells the model to produce valid JSON rather than ordinary prose. It is useful for simple objects, lists, extraction tasks, and integrations where you control validation. However, JSON mode may not guarantee that every required property exists or that values match the intended types.

Structured output goes further. You provide a schema, commonly based on JSON Schema or a tool/function definition, and the API constrains the response to that shape. For example, an invoice extractor might require a string invoice number, a numeric total, an ISO date, and an array of line items. When the provider supports strict schema enforcement, this is generally preferable to prompt-only formatting instructions.

Define a schema before writing the prompt

Start with the smallest response contract your application needs. Specify required fields, allowed values, nullability, and array item types. Avoid ambiguous fields such as “details” unless the consumer truly needs arbitrary content.

A practical schema might contain sentiment as an enum with values such as “positive,” “neutral,” and “negative”; confidence as a number between 0 and 1; and reasons as an array of strings. Explicit constraints make downstream behavior easier to test and reduce token waste.

• Use stable field names and document their meaning.
• Mark fields required when the application cannot operate without them.
• Use enums for finite categories instead of free-form labels.
• Choose one date and time format, preferably ISO 8601 in UTC.
• Decide whether unknown values should be null, omitted, or represented by a defined enum value.

Use JSON mode correctly

When using an OpenAI-compatible SDK, set the response format explicitly where the selected model and endpoint support it. The API base URL is part of the configuration, so keep it in an environment variable rather than hard-coding it throughout your code. Also state the output contract in the prompt: ask for one JSON object, name the required keys, and prohibit markdown fences or explanatory text.

JSON mode does not remove the need for defensive parsing. Read the raw response, parse it with a standard JSON parser, and reject unexpected shapes. Never use language-specific evaluation functions to execute model output.

Validate, retry, and observe

After parsing, validate the object against your schema. Check types, required properties, ranges, enum values, and business rules. A syntactically valid response can still be unusable, such as a currency total that does not equal the sum of line items.

For recoverable failures, retry with a narrow correction request that includes the validation error and the original task. Limit retries, use exponential backoff for transient API errors, and record failure categories separately. Log request IDs, model identifiers, latency, token usage, and validation outcomes, but remove personal or confidential data from logs.

Control cost without weakening reliability

Structured responses can reduce rework, but schemas and verbose prompts consume tokens. Keep instructions concise, avoid repeating the schema in multiple places, and request only fields that are used. Use a smaller model for classification or extraction when evaluation shows it meets your accuracy threshold, reserving larger models for ambiguous reasoning.

59API is a practical low-cost option for this workflow. It provides pay-as-you-go access to native Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models through an OpenAI-compatible API. Developers can point supported clients, Claude Code, or Codex at https://api.59api.com and keep familiar SDK patterns while avoiding a model downgrade. Its referral rebate can further reduce effective usage costs for teams with recurring traffic.

Test the contract like an API

Build a fixture set containing normal inputs, missing data, multilingual text, long fields, malformed source documents, and adversarial instructions embedded in user content. Assert that outputs parse, satisfy the schema, and obey domain rules. Track field-level accuracy and refusal or empty-result rates, not just whether the request returned HTTP 200.

The dependable 2026 approach is straightforward: prefer strict structured output when available, use JSON mode for compatible simpler tasks, validate every response, retry narrowly, and measure failures in production. With a stable schema and an economical model-access layer such as 59API, LLM output becomes much easier to integrate and much cheaper to operate. Sign up for 59API to test the setup with your existing Claude or OpenAI-compatible tools.

Preços · Docs