OpenAI drops a new toolkit for building AI agents

For the past year, stitching together a production-ready AI agent with OpenAI's APIs has been a messy chore. Developers had to juggle between Chat Completions and the Assistants API, write endless boilerplate glue code to orchestrate prompts, and pray that multi-turn tool calls didn't fall apart invisibly.

OpenAI rolled out a comprehensive developer stack to fix that: the new Responses API, an open Agents SDK, integrated observability tools, and three major built-in tools covering web search, file search, and computer use.

Meet the Responses API

The centerpiece of the release is the Responses API. OpenAI describes it as a new primitive that combines the simplicity of Chat Completions with the built-in tool abilities of the Assistants API. Instead of forcing developers to loop requests manually, a single Responses API call can now orchestrate multiple tools and model turns to complete complex, multi-step tasks.

Alongside the API, OpenAI is integrating three turnkey tools directly into the platform:

Web search gives agents real-time browsing capabilities backed by the same tech behind ChatGPT search. Available for gpt-4o and gpt-4o-mini, it scores 90% and 88% respectively on the SimpleQA benchmark, complete with inline citations. Search queries are priced at $30 per thousand for GPT-4o search and $25 per thousand for 4o-mini search.

File search handles retrieval over massive document sets, featuring built-in query optimization, metadata filtering, and reranking. It costs $2.50 per thousand queries, with storage at $0.10 per GB per day after the first free gigabyte.

Computer use brings the Computer-Using Agent (CUA) model that powers OpenAI's Operator straight to developers. It captures mouse clicks and keystrokes so machines can navigate software and web interfaces. On benchmarks, the model posted a 38.1% success rate on OSWorld, 58.1% on WebArena, and 87% on WebVoyager.

The sunset of Assistants API

With Responses API taking center stage, OpenAI is cleaning up its fragmented developer lineup.

Chat Completions isn't going anywhere and will still receive model updates for simple, tool-free tasks. However, the Assistants API is officially on borrowed time. Once the Responses API gains full feature parity—adding code interpreter, thread objects, and assistant primitives—OpenAI plans to formally sunset the Assistants API around mid-2026, accompanied by a migration path for existing data.

To help developers monitor workflows, OpenAI also bundled native tracing tools into the platform, letting builders inspect agent steps and debug edge cases without duct-taping third-party telemetry.

Why it matters

This release turns AI agents from a messy research demo into a standardized software engineering primitive.

For builders, the barrier to creating functional agents just dropped dramatically. Instead of spinning up external vector databases, search scraping APIs, and bloated orchestration libraries, developers can now deploy an agent with web browsing, document retrieval, or GUI control in just a few lines of code. If you are building customer support bots, research copilots, or enterprise search assistants, it is definitely worth testing right away.

For the broader industry, OpenAI is moving aggressively down the stack. While Anthropic has long advocated for lean agent workflows and pioneered computer use, OpenAI is wrapping model inference, tool execution, and observability into a tight, proprietary ecosystem. That puts immense pressure on standalone framework makers and middleware startups whose main value was managing these exact workflows.

We think this is a smart, necessary shift that eliminates the clunky developer experience of the Assistants API. However, the catch is obvious: costs can escalate quickly, especially with search queries at up to $30 per thousand, and GUI automation remains fragile with success rates below 40% on complex desktop benchmarks. It makes building easy, but locks your architecture squarely into OpenAI's walled garden.