For the past few years we have been interacting with AI that can answer questions, write documents, generate code, and search the web.
Computer use, which allows AI to operate software, navigate a browser, and create or edit files, has been pretty underwhelming in my experience. The results often needed so much fixing or improving that I might as well have done the work myself. However, OpenAI’s GPT-6 Astra and the recently released Claude Opus 5.5 have started to change that.
AI can now use your computer with great results.
OpenAI’s new Astra model is particularly good at this. Instead of simply telling you how to do something, it can look at what is happening on your screen, click buttons, navigate menus, open files, type, and interact with software much like a person would.
OpenAI describes Astra as its strongest model yet for computer use, including tasks such as updating records, working with documents, installing and testing software, troubleshooting problems on screen, and moving through multi-step workflows across different applications.
I recently tested this myself with video editing.
I pointed Astra to a folder containing several video clips and asked it to use a video editing application. I gave it instructions about how I wanted the video edited, including cuts and captions, and then let it work.
It was surprisingly good.
It opened the software, imported the clips, navigated the interface, made the edits, and assembled the video. I still wanted to make a few small adjustments afterward, but most of the work was already done.
What was particularly interesting was that Astra had never been specifically trained to use this application. It was looking at the interface and figuring it out. That is a pretty significant shift.
Anthropic is moving in the same direction. Claude’s computer-use capabilities can interpret what is on the screen, move the cursor, click, type, open files, use a browser, and interact with development tools. Anthropic has also integrated computer use into Claude Cowork and Claude Code.
There are obviously some important questions that come with giving AI this level of access. An agent that can click, type, open files, and interact with accounts can also make mistakes. That means permissions, privacy, and human oversight become very important.
However, if you are cautious and creative you can use this feature as a great productivity tool. For years we have been asking: What can AI create for me? We are quickly moving toward a different question: What can AI actually do for me on my computer?