Responses API

Overview

The Responses API is OpenAI’s agent-native API primitive, launched in March 2025 as the successor to both the Chat Completions API and the Assistants API. It provides a unified endpoint (POST /v1/responses) that combines text generation, built-in tool execution, multi-turn state management, and multimodal I/O into a single request/response cycle with an automatic agentic loop.

Timeline

Mar 2025: Responses API launched with web search, file search, and computer use tools
Aug 2025: Assistants API deprecated (sunset Aug 2026); migration guidance published
2026-04-02-AI-Digest — Shell tool released for enhanced programmatic interaction
2026-04-05-AI-Digest — Extended with shell tool, agent execution loop, hosted container workspaces, context compaction, and reusable agent skills
2026-04-07-AI-Digest — Full agentic platform with hosted shells, context compaction, and reusable agent skills

Key Features

Agentic loop: Model autonomously calls multiple tools within a single API request
Built-in tools: Web search, file search, code interpreter, image generation, computer use, remote MCP servers
previous_response_id: Server-managed conversation state with forking support
Background mode: Long-running tasks without holding connections, with webhook notifications
Streaming: Typed SSE events with sequence numbers
Strict function calling: Functions are strict by default (unlike Chat Completions)

Built-in Tools

Tool	Description
`web_search_preview`	Internet search with sourced citations
`file_search`	Semantic + keyword search over uploaded files in vector stores
`code_interpreter`	Sandboxed Python execution in managed containers
`image_generation`	GPT Image model integration with streaming previews
`computer_use_preview`	Desktop automation via mouse/keyboard actions
`mcp`	Connect to remote MCP servers (Streamable HTTP / HTTP-SSE)
`function`	Custom function definitions (strict schema by default)

Comparison with Anthropic

The fundamental architectural difference is that OpenAI’s Responses API runs an agentic tool loop server-side (one API call returns the final answer), while Anthropic’s Messages API uses a client-side tool loop (developer executes tools and sends results back). This gives OpenAI simplicity for built-in tools but gives Anthropic developers full control over execution.

OpenAI — Parent company and platform
MCP — Model Context Protocol, supported natively in Responses API
Claude Code — Anthropic’s competing agentic developer tool
Cursor — IDE integration with Responses API