This post was created with AI assistance and reviewed for accuracy before publishing.
On March 11, 2025, OpenAI released a set of tools for building AI agents: the Responses API, built-in web search and file search, a computer use tool, and an open-source Agents SDK. The announcement is the primary source for what follows. No invented features or timelines.
What the Responses API is
The Responses API is a new API primitive that combines the simplicity of the Chat Completions API with the tool-use capabilities of the Assistants API. According to OpenAI, developers can solve increasingly complex tasks with a single API call using multiple tools and model turns. It supports a unified item-based design, streaming events, and SDK helpers like response.output_text. The API is available to all developers and uses standard token and tool pricing; there are no separate fees for the API itself.
The Assistants API will be deprecated, with a target sunset in mid-2026. OpenAI plans a migration guide to move data and applications to the Responses API. Until deprecation is formally announced, new models will still ship to the Assistants API.
Built-in tools
Web search: Available with gpt-4o and gpt-4o-mini in the Responses API. Returns up-to-date answers with citations. OpenAI reports 90% and 88% accuracy on SimpleQA for the search preview models. Pricing: $30 and $25 per thousand queries respectively. Also exposed via gpt-4o-search-preview and gpt-4o-mini-search-preview in Chat Completions.
File search: Search over large document sets with support for multiple file types, query optimization, and metadata filtering. Priced at $2.50 per thousand queries and $0.10/GB/day for storage (first GB free). Intended for RAG-style pipelines, customer support, and knowledge bases.
Computer use: A research preview tool powered by the same model behind Operator. Generates mouse and keyboard actions for browser automation. Available to developers in usage tiers 3–5. OpenAI states it achieves 38.1% on OSWorld (a benchmark for real-world computer tasks), 58.1% on WebArena, and 87% on WebVoyager. The post explicitly notes the model is "not yet highly reliable for automating tasks on operating systems" and recommends human oversight. Pricing: $3/1M input tokens, $12/1M output tokens.
The Agents SDK
The open-source Agents SDK helps orchestrate single-agent and multi-agent workflows. It offers tracing and observability, configurable guardrails for input and output validation, and handoffs between agents. It works with the Responses API and Chat Completions API, and is designed to support other providers' models that expose a Chat Completions-style endpoint. Python support is available now; Node.js is planned. The docs include quickstart and examples.
What developers can actually build
OpenAI gives examples in the announcement: shopping and research assistants using web search, customer support agents using file search over docs, and browser automation for QA and data entry using computer use. Coinbase and Box are cited as early users of the Agents SDK. The computer use tool is framed as suitable for browser workflows; for local OS automation, reliability is limited and oversight is advised.
How to get started
The Responses API quickstart and Agents SDK docs are the canonical references. The Playground supports web search and file search for experimentation. If you are on the Assistants API, you can keep using it for now, but the migration path runs through the Responses API.
Top comments (0)