Published March 9, 2026, OpenAI's one-year retrospective on the Responses API argues that the endpoint has become the default surface for building agent-native products rather than a mere successor to Chat Completions. Where Chat Completions treats each call as a stateless turn, the Responses API (documented at /v1/responses) handles conversation state server-side, supports reasoning controls and summaries, orchestrates built-in tools during reasoning, and runs long jobs through Background mode — the primitives teams kept rebuilding by hand before it existed.
The post is structured around developer stories that show the API in production. Raindrop AI built a monitoring platform that detects agent failures using background-analysis workflows, leaning on the API's ability to run long, asynchronous inspection passes without holding a client connection open. Repo Prompt built a deep-reasoning system that deliberately separates the context-gathering phase from the analysis phase, using reasoning controls to spend compute where it matters. These are the kinds of multi-phase pipelines the stateful API is designed to make routine.
Other examples highlight custom tool integration. Collxn built a vinyl-collection interface wired to 17 custom Discogs API tools, letting the model query and cross-reference a music catalog conversationally. Arcade used computer-use analysis to turn screen recordings into product demos, and Hexagon built a brand-visibility tracker with a multi-agent content-generation pipeline. Across these, the common thread is that the API's built-in tools — web search, file search, computer use, and shell tools — compose with developer-defined tools instead of forcing a choice between them.
The models named in the retrospective span the current lineup: GPT-5.4 for general-purpose and frontend-heavy work, GPT-5.3-Codex for coding agents, GPT-5.2, GPT-5 nano for cost-sensitive high-volume steps, and computer-use-preview for automation loops. The point OpenAI stresses is that these models were designed for the Responses API surface — reasoning summaries, tool calls emitted during reasoning, and persisted conversation state all assume the stateful endpoint rather than the legacy stateless one.
Statefulness is the throughline. Conversation state and the Conversations API give durable, replayable threads, so an agent's history — including its reasoning and tool calls — survives across turns and can be inspected after the fact. Combined with Background mode for long-running responses and webhooks for event-driven completion, this lets teams build workflows that run for minutes or longer without bespoke queueing and polling infrastructure. Observability, OpenAI notes, becomes a property of the platform rather than something each team bolts on.
The retrospective also positions the Responses API as the foundation for higher-level tooling. Codex is described running with an App Server plus MCP, and the broader AgentKit and Agents SDK stack sits on top of the same primitives, so a team can start with raw Responses calls and graduate to orchestrated multi-agent systems without switching APIs. Support for Skills and server-side compaction, added earlier in 2026, further extends how much of a long-running agent's context and behavior the platform manages directly.
On adoption, OpenAI says thousands of developers are now building with the Responses API across customer support, legal, life sciences, and travel — a spread meant to show the endpoint is not just for coding assistants but for regulated, document-heavy, and workflow-oriented domains. The company frames the first year as validation that agent-native primitives, not prompt engineering alone, are what let teams move from prompts to products.
The strategic message is unambiguous: Chat Completions is now the legacy path, and new builds should start on the Responses API to get stateful conversations, background execution, reasoning controls, and the built-in tool ecosystem without reinventing the glue. For developers modernizing an older integration, the retrospective effectively serves as a migration nudge — the newest models, tools, and platform features are landing on /v1/responses first.