Alibaba's OpenGCUA Opens Computer Use Stack

OpenGCUA's important idea is architectural: computer use becomes a branching, feedback-driven program instead of a slow screenshot-click loop. The benchmark gains are promising, but the preview release and self-reported results need independent replication.

• –Code execution lets agents loop, branch, inspect intermediate results, and batch repetitive work.
• –Harness-to-model decision transfer gives models more control over when to observe and act.
• –Open trajectories across eight models create useful material for studying failure recovery and wasted steps.
• –The cross-device SDK connects computers, browsers, and phones, but remains preliminary and permission-heavy.
• –OpenGCUA wraps existing models rather than introducing a new one, potentially making cheaper models more practical for computer-use tasks.