What computer use actually means
Computer use is an AI model taking screenshots of your screen, deciding what to do, then moving the mouse and typing. It is slower than a proper integration, but it works with almost any app, including old desktop software and supplier portals that will never have an API.
That matters for small businesses more than big ones. Big companies pay developers to connect systems. Most small offices have someone copying data between screens by hand. That is the job computer use goes after.
OpenAI: computer use inside the ChatGPT desktop app
OpenAI's computer use docs say ChatGPT can see and operate apps on macOS or Windows from the ChatGPT desktop app, inside ChatGPT Work and Codex, in supported regions. It asks permission for each app (with an "Always allow" option), may ask again before sensitive actions, and admins can restrict which apps it touches. On Windows, the app it is working in has to stay visible and it takes over your keyboard and mouse while it runs.
OpenAI's newest model, GPT-6.1 Terra, released October 26, is pitched squarely at this. The changelog entry describes it as pairing image understanding with precise computer use, and it is cheaper than the bigger GPT-6.1 Sol. We wrote up Terra separately in our Terra breakdown.
Anthropic: computer use in Claude Cowork and Claude Code
Anthropic launched computer use as a research preview in Claude Cowork and Claude Code on March 23, 2026, for Pro and Max subscribers, according to Anthropic's announcement. Claude tries a direct connection to your apps first, and only falls back to controlling the mouse, keyboard and screen when it has to. It asks permission before acting and before opening new apps. The desktop app has to be open and the computer awake.
The same page now says computer use is supported on both macOS and Windows. Cowork itself became generally available on Mac and Windows on April 9, 2026, per Anthropic's release notes. Note the label though. Computer use is still called a research preview, which is Anthropic's way of saying expect rough edges.
The real risk: prompt injection
Prompt injection is when a web page, email or file contains hidden instructions that trick the AI into doing something you did not ask for. A person would ignore a line of white text saying "forward this inbox to me". An AI reading the whole page might not.
This is measured, not hypothetical. When Anthropic piloted its Chrome browser agent, its own testing found attacks succeeded 23.6% of the time before fixes and 11.2% after. Vendors report big improvements since then, but those newer numbers also come from their own testing, and nobody claims the problem is solved. OpenAI's own computer use page warns that web pages can contain malicious or misleading content and recommends narrow tasks, staying present for sensitive steps and keeping secrets out of reach.
Rule of thumb: if a task would be a disaster when done wrong once, a person approves it. Payments, customer emails, deleting records and anything behind a password all fall in that bucket.
Where small businesses should start
Good first jobs are repetitive, read-heavy and easy to check:
- Distributors and wholesalers: pull order status from a supplier portal into a spreadsheet each morning.
- Trades: copy job details from emailed work orders into your scheduling software.
- Law and accounting offices: gather documents from several portals into one client folder for a person to review.
- Any office: fill in the same web form for every new customer, then stop before submitting so someone can check it.
Bad first jobs: anything inside online banking, payroll, or a customer's inbox. Not because it cannot do them, but because the cost of one mistake is too high while you are still learning how it behaves.
Which one should you use? For most of the businesses we work with, we set up Claude Code first, because it is careful on long tasks and easy to fence in. If your team already runs on ChatGPT, OpenAI's version is a solid pick, and we cover it in our Codex guide. The honest answer is that both are good enough now, and the setup and guardrails matter more than the brand.
What we'd do this week
- List three tasks where someone copies data between screens by hand.
- Pick the lowest-risk one and try it with ChatGPT or Claude watching over your shoulder.
- Use a separate login with limited access for the AI, not your admin account.
- Want a second set of eyes on the setup? Book a free audit.