Why we built the Responses API | OpenAI Developers

Introduces the `/v1/responses` API as a successor to `/v1/chat/completions` for agentic workflows. Key technical contributions include: preserving the model's reasoning state across turns (improving TAUBench by 5% and cache utilization), emitting polymorphic items (tool calls, structured outputs, intermediate reasoning steps) alongside messages, integrating hosted tools (File Search, Image Gen, Code Interpreter, MCP) for server-side execution, and providing a more robust and efficient API design for multimodal and reasoning models. The API is designed to be stateful, multimodal, and efficient, with internal benchmarks showing 40-80% better cache utilization.