Structured Outputs Are Now GA on the Claude API

Structured outputs — the ability to enforce a JSON schema on Claude's response — moved to general availability on the Claude API. This removes the need for prompt-based JSON coercion (asking Claude to "always return valid JSON") and replaces it with guaranteed schema enforcement at the API level.

How It Works

Add the betas parameter is no longer needed — the feature is now stable and enabled via the standard API. Pass a JSON schema in the tool definition or use the dedicated response format parameter. Claude will always produce output that validates against your schema.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-opus-4-5",
  max_tokens: 1024,
  tools: [{
    name: "extract_data",
    description: "Extract structured data from text",
    input_schema: {
      type: "object",
      properties: {
        name: { type: "string" },
        confidence: { type: "number", minimum: 0, maximum: 1 }
      },
      required: ["name", "confidence"]
    }
  }],
  tool_choice: { type: "tool", name: "extract_data" },
  messages: [{ role: "user", content: "Extract data from: ..." }]
});

Why This Matters for Architects

Before structured outputs, getting reliable JSON from Claude required extensive prompt engineering — examples, explicit format instructions, post-processing validation. Every approach had failure modes at the edges, especially when Claude was uncertain about how to represent a value in the schema.

Structured outputs eliminate this class of problem entirely. The model cannot produce output that violates the schema. This changes how you architect data pipelines: you no longer need a validation/retry loop for JSON extraction tasks.

What Changes for Prompt Engineering

The CCA-F Domain 4 question this raises: when you have structured outputs available, is the correct answer still to add few-shot examples? No — structured outputs make few-shot formatting examples unnecessary for JSON extraction. The exam tests whether you know the difference between format enforcement (schema) and content guidance (few-shot examples).

Use structured outputs for: data extraction, classification tasks, API response generation. Continue using few-shot examples for: style guidance, reasoning patterns, domain-specific decision criteria — things that aren't about output format.

Limitations

Structured outputs constrain the response structure, not the content quality. Claude can still produce factually incorrect values that happen to be schema-valid. Structured outputs solve the parsing problem; they don't solve the accuracy problem. For high-stakes data extraction, you still need validation logic on the values themselves.