When you ask a chatbot a question, it answers with text. Computer use is different: it lets an AI model see a screen and operate a mouse and keyboard directly, clicking buttons, filling in forms, and switching between apps the way a person would — without anyone building a custom integration for each piece of software first.
What It Is
Computer use is an AI capability, not one single product. Anthropic introduced it in October 2024 as a feature of Claude, and OpenAI shipped its own version — a model it called a Computer-Using Agent, or CUA — inside Operator, a research preview launched in January 2025. Both work the same basic way: the model takes a screenshot, decides what to do next, and then moves the cursor, clicks, or types on the same graphical user interface a human uses, rather than through a programming interface built specifically for AI.
That distinction matters because most software has no AI-friendly way in. A hospital’s scheduling system, a decades-old accounting tool, or an internal company dashboard may never get an official API. Computer use lets a model operate that software anyway — the way a new employee would, by looking at the screen and clicking around.
How It Works
The model runs in a loop. It receives a screenshot of the current screen, reasons about what needs to happen next based on the task it was given and everything it has done so far, and picks an action: move the cursor to a coordinate, click, type text, scroll, or take another screenshot to check the result. It repeats this cycle — one screenshot, one action — until the task is finished or it needs a person to step in.
Getting this right proved to be a genuinely hard engineering problem. Early testing showed models struggled just to judge how many pixels to move a cursor to land on the right button, a skill humans do without thinking. Anthropic’s version scored 14.9% on the OSWorld benchmark of real desktop tasks — ahead of other models at the time (7.7%) but far below a human’s typical 70–75%. OpenAI’s CUA showed a similar gap in its own testing: around 38% on OS-level tasks and 58% on web tasks. Those numbers describe the current state of the technology rather than a finished product: computer use is genuinely useful today, but slower and less reliable than either API-based automation or a person doing the same task.
Why It Matters
Computer use unlocks a different kind of automation than a chatbot answering questions, or an AI agent calling a documented API. It works with software exactly as it already exists, without waiting for a vendor to build an integration — which is why it sits behind products marketed as “agents that get work done”: testing an app the way a user would, filling out forms across systems that don’t talk to each other, or working through a multi-step task on an old internal tool.
That flexibility is also the tradeoff. Because the model reacts to whatever appears on screen, it can be misled by what it sees there. A malicious pop-up, a hidden instruction buried in a web page, or altered text on an otherwise ordinary-looking form can steer it off task — a screen-based variant of prompt injection aimed at what’s rendered on a page rather than what’s typed into a chat box. Anthropic classifies computer use at its second-lowest AI Safety Level and recommends running it inside an isolated virtual machine with no access to sensitive accounts, plus human confirmation before any consequential step — entering payment details, sending money, or agreeing to a contract.
OpenAI’s version of the idea has already changed shape once: Operator launched as a standalone product in January 2025, was folded into ChatGPT’s built-in “agent mode” in July 2025, and was shut down as a separate app the following month. The underlying CUA model didn’t disappear — it now runs inside ChatGPT and OpenAI’s developer tools — but the switch is a reminder that this is still an early, fast-moving capability rather than a settled product category.
How to Try It
Computer use is a developer-facing feature today, reached through the Claude API rather than a standalone consumer app. Anthropic publishes a reference implementation with a ready-to-run sandbox, and ChatGPT users can reach the same underlying capability through “agent mode” in the regular ChatGPT interface. There’s no separate subscription fee for computer use itself — it’s billed like any other API call, at standard per-token rates (as of July 2026, Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, per Anthropic’s pricing page), plus a small addition for the tool definition and for each screenshot the model reviews.