Alibaba researchers open a computer-use agent setup that helps eight models

An Alibaba research team released OpenGCUA, an open setup for AI agents that operate a computer by mixing clicks with code, shell commands and app interfaces, instead of only clicking on screenshots. On the OSWorld-v2 benchmark, it raised scores for all eight models they tested, including GPT-6 Astra and Claude Opus 5, by 19.6 points on average while taking 53% fewer steps than click-only runs. They published the code, prompts and test runs, plus early versions of a harness for running on a real computer and a kit for running one task across a computer, phone and browser. These are the authors' own results on one benchmark plus four real Mac tasks, so treat it as an early research release. Sorted take: the useful lesson is the design, not the score. Agents that can run code or call an app directly do better than ones that only click, so give your automation those routes first.