OpenAI Structured Outputs

• Profile: openai.so.2026-04-30
• API: Responses API text.format.type = "json_schema" and Chat Completions response_format.type = "json_schema" with strict: true
• Model behavior: Strict — rejects schemas with unsupported keywords

Supported Keywords

OpenAI supports a subset of JSON Schema. The following keywords are rejected for gpt-4o-2024-08-06 / gpt-4o-mini Structured Outputs:

• allOf, oneOf, and root-level anyOf
• not, if/then/else
• dependentRequired, dependentSchemas
• propertyNames, maxProperties, minProperties, uniqueItems, contains, unevaluatedProperties

Nested anyOf remains supported when each branch is valid for the Structured Outputs subset. patternProperties works for the base Structured Outputs model surface, but OpenAI documents additional restrictions for fine-tuned models.

Value Restrictions

• additionalProperties: Must be false (no object schemas)
• format: Only ["date-time", "time", "date", "duration", "email", "hostname", "ipv4", "ipv6", "uuid"]

Structural Limits

Limit	Value
Root schema type	Must be object
Root composition	Must not use anyOf, oneOf, allOf, enum, or not
additionalProperties	Must be false on all objects
required	Must include every property on every object
Array schemas	Must declare items
Max nesting depth	10
Max total properties	5,000
Max total enum values	1,000
Max string length budget	120,000
Enum string budget	15,000 characters when an enum has more than 250 values
External $ref	Not supported

The depth limit is evidence-based: the retained probe_limits_2026-06-16.json records inline object depth 10 as accepted and depth 11 as rejected by the Responses API. That probe does not contain local $ref chains, so SchemaLint does not reinterpret it as proof of a separate semantic reference-depth rule. Other budgets in this table follow the linked provider contract and are covered by offline profile/conformance tests; live provider probes remain manual and credential-gated.

Reference

OpenAI Structured Outputs documentation