OpenAI Ships Computer Use in the Responses API

OpenAI launched the Responses API and bundled three built-in tools with it: web search, file search, and computer use. Computer use is the one that lets a model drive a computer the way a person does. It looks at a screenshot, decides on an action — click, type, scroll, drag, keypress — with pixel coordinates, that action is executed, a fresh screenshot goes back, and the loop repeats until the task is finished. The model snapshot released on this date is computer-use-preview-2025-03-11. It is the same underlying capability behind Operator, the consumer product OpenAI had launched on 23 January 2025; this release put it behind a developer API. Two things about access are worth getting right, because they are commonly reported wrongly. It was not enterprise-only: it opened to developers in API usage tiers 3 to 5. And it is billed in tokens like any other model — $3 per million input tokens and $12 per million output — plus a fee for each tool call, not per task step. Executing the actions is the developer's job. OpenAI publishes a sample app showing how to wire the loop up to a browser or a desktop, but the API itself does not ship an automation layer. OpenAI's own guidance is blunt about the risks: run it in an isolated browser or virtual machine with an allow list of sites and actions, treat whatever appears on screen as untrusted because text in a page cannot grant the model permission, ask the user before anything consequential, and put hard limits on steps, time and cost with a way to cancel.