This piece assumes Responses/Agents context — it covers computer use and built-in tools only. Full series in navigation below.
OpenAI’s agent story is incomplete without the action surface. On March 11, 2025, alongside Responses and the Agents SDK, OpenAI shipped built-in tools — including web search, file search, and computer use — so models can retrieve, browse, and operate software UIs instead of only emitting text.
Confirmed tool layers
- Web search / file search: Hosted retrieval tools on Responses so developers do not have to assemble every RAG or browse pipeline from scratch.
- Computer use: Models inspect screenshots and return structured UI actions (click, type, scroll, screenshot requests) for your harness to execute — or drive custom Playwright/Selenium/VNC/MCP harnesses via ordinary tool calling.
- Code-execution harness path: OpenAI documents models trained to mix visual UI interaction with short scripts / DOM-style programmatic control inside a runtime.
- Model evolution: OpenAI’s current computer-use docs highlight newer frontier models (including gpt-5.4-class training notes) that improve this screenshot-reason-act pattern and keep the same Responses computer tool shape.
What this is — and is not
Confirmed: computer use is a first-party Responses tool pattern for UI automation; developers remain responsible for executing actions in an environment they control. Not confirmed here: any claim that OpenAI guarantees safe unsupervised production automation across arbitrary enterprise desktops — treat deployment risk, approval gates, and sandboxing as engineering requirements, not marketing footnotes.
Where it sits in the stack
ChatGPT is the consumer conversation surface. The API model family is the intelligence SKU. Responses + SDK are the integration and orchestration layers. Computer use and hosted tools are how those layers touch the world — browsers, files, and GUIs — which is why agent incidents and eval breakouts matter as much as SWE-bench scores in 2026 coverage.
Our take
In our view, “computer use” is the moment OpenAI’s product stack stops being a chat vendor and becomes an automation platform. The interesting competition is no longer only who has the smartest Sol/Mythos twin — it is who ships the safest, steerable action loop.
Series: OpenAI Product Stack · next: consequence file.
