Computer-use agent claims $0.0002 a step

A new macOS computer-use tool is making a striking cost claim. It combines OCR, which reads text from the screen, with action classification to decide what to click or type next. Its stated cost is roughly $0.0002 per step, far below the pricing often associated with vision-model agents that process screenshots directly.

If that holds up in real work, it matters because cost is one of the main limits on automating longer desktop tasks. A task that takes dozens or hundreds of screen actions can become expensive quickly when every action requires a full vision-model call. A lower-cost approach could make it more practical to test automation for repetitive internal work, especially where staff move between browser tabs, spreadsheets and desktop software.

The important word here is claims. Low cost alone does not tell you whether the agent handles messy screens, changing layouts, exceptions or sensitive business systems reliably. Those are the things operators should test before putting it near a live workflow. Still, the early attention is notable. The project reached 156 GitHub stars, at a reported rate of 147 per day, just one day after it was posted. You can see the project details in the original repository.

For a business owner, the practical takeaway is to watch for computer-use tools that split screen reading from decision-making, rather than treating every screen interaction as an expensive image-analysis problem. This is the kind of thing we build into an AI command centre, where cost, reliability and human review need to be visible before an automation is trusted.