TopolearnPrograms need data they can trust to parse, not prose that is usually JSON.
Asking "reply in JSON" works most of the time, which is not good enough for a program. Structured outputs make Claude's reply follow a JSON Schema you supply, so the result always parses and has the fields you asked for.
There are two ways to use it in Python:
client.messages.parse(..., output_format=MyModel): pass a Pydantic model (a Python class that declares typed fields and validates data). The SDK builds the schema, sends it, and gives you a ready object inresponse.parsed_output.client.messages.create(..., output_config={"format": {"type": "json_schema", "schema": {...}}}): send a schema yourself andjson.loadsthe text block.
The feature is generally available. Older code used a top-level output_format request field with a beta header; that field is deprecated in favour of output_config.format.
For tools, the same guarantee comes from "strict": True on a tool definition, which makes the tool inputs match the schema.
Limits to know: schemas must set additionalProperties to false, and some JSON Schema features (number ranges, string lengths, recursive schemas) are not enforced. The first request with a new schema is a bit slower while it is compiled. A refusal or hitting max_tokens can still produce output that does not match, so check stop_reason first.
This also replaces an old trick: starting the assistant's answer with { (prefill), which current models reject.
Example
import anthropic
from pydantic import BaseModel
class Contact(BaseModel):
name: str
email: str
wants_demo: bool
client = anthropic.Anthropic()
resp = client.messages.parse(
model="claude-opus-5-5",
max_tokens=512,
messages=[{"role": "user", "content":
"Extract: Jo Park (jo@example.com) asked us to book a demo next week."}],
output_format=Contact,
)
contact = resp.parsed_output
print(contact.name, contact.email, contact.wants_demo)
# Jo Park jo@example.com TrueRemember
messages.parse(output_format=Model)returns a validated object- Raw form:
output_config={"format": {"type": "json_schema", ...}} strict: Trueon a tool guarantees schema-valid tool inputs
Watch out for
- Using the deprecated top-level output_format request field from older examples.