We Built the Responses API (developers.openai.com)

OpenAI announced the Responses API, a purpose-built interface for GPT-5 that unifies chat, agentic tool use, and multimodal inputs into a stateful “reasoning and acting” loop. Unlike the turn-based Chat Completions endpoint, Responses preserves the model’s internal reasoning state across steps (so intermediate thought processes survive between calls) while keeping that chain-of-thought encrypted and hidden from clients. That lets models investigate, call hosted tools, and report structured results—think of a detective that keeps private notes but returns receipts for actions—making it far better suited to complex, multi-step workflows and agentic applications. Technically, Responses supports first-class multimodality (text, image, audio, function/tool calls), emits multiple output items (final answers, tool calls, intermediate steps), and exposes conveniences in the SDK (semantic streaming events, output_text helpers, organized multimodal params). Hosted server-side tools (file_search, code_interpreter, web search, image gen, MCP) reduce developer plumbing and latency. OpenAI reports concrete gains: +5% on TAUBench with GPT-5 and 40–80% better cache utilization versus Chat Completions, translating to lower latency and costs. Chat Completions will remain supported, but Responses is positioned as the default for persistence, safety-aware reasoning, and richer agentic/multimodal apps.