Structured Outputs (OpenAI)
OpenAI's Structured Outputs feature, introduced August 2024 with gpt-4o-2024-08-06, lets developers supply a {{JSON Schema}} via the response_format parameter with strict: true. The model's decoder is gated server-side so output is guaranteed to match the schema. On OpenAI's internal eval, the feature scored 100% schema adherence vs under 40% for prompt-only GPT-4.
Structured Outputs is OpenAI's server-side constrained decoding implementation, released August 6, 2024 alongside the gpt-4o-2024-08-06 model. Developers pass a JSON Schema (or a Pydantic / Zod class via the SDKs) under the response_format parameter with strict: true; the inference server enforces the grammar at token-mask level so the returned string is always parseable JSON conforming to the schema. The feature is exposed in two forms: as a response_format on regular completions, and via strict: true on tool definitions for Function Calling (LLM). The Python and Node SDKs auto-convert Pydantic / Zod types to JSON Schema and deserialize on the return trip. OpenAI reported 100% adherence on their complex-schema eval vs under 40% for the older gpt-4-0613 with prompt-only formatting instructions. Limitations: strict mode supports only a subset of JSON Schema (no unbounded recursion, limited use of anyOf, no top-level non-object schemas). Schema compilation adds a small first-call latency that is cached on repeat. The feature addresses the syntactic side of Format-Following Failures in LLMs but does not validate semantic correctness of extracted values.