Claude 3.5 “Computer Use” Explained: See the Screen, Click, and Type to Get Tasks Done

Claude 3.5’s most interesting update is that it pushes AI from “can chat” to “can operate.” In the public beta, Claude 3.5 can view what’s on your screen, move the cursor, click buttons, and type into input fields to complete step-based tasks. Below is an editor’s breakdown of what’s changed, so you can judge whether it’s worth jumping in immediately.

What Claude 3.5’s new “Computer Use” can do

“Computer use” means you give Claude 3.5 a goal, and it executes it like a person would by following the on-screen UI flow: first it looks at the screen, then decides where to click and what to type. It’s well-suited for work that involves many steps, such as filling in items one by one in a web admin panel, or organizing information from Page A into a form on Page B. Anthropic also makes it clear this is still an experimental capability—Claude 3.5 may occasionally misclick or miss a step, so you’ll need to correct it as you go.

Availability: API access with multi-platform support

Right now, Claude 3.5 computer use is offered as a public beta via API. Developers can build with it directly through the Anthropic API. At the same time, Claude 3.5 is also available on Amazon Bedrock and Google Cloud’s Vertex AI, making it easier for companies to integrate it into existing cloud setups. For teams, this means Claude 3.5 isn’t just a demo—it’s the kind of capability that can be brought into workflow systems for automation.

Model updates: Sonnet upgrade and Haiku addition

This release includes an upgraded Claude 3.5 Sonnet (on AWS you may also see it labeled Sonnet V2), with a focus on improving coding and tool-use performance while keeping similar cost and speed. The other track is Claude 3.5 Haiku, designed for faster responses and lower latency—better suited for user-facing products and smaller sub-tasks. For selection: if you need more reliable execution on complex workflows, prioritize Claude 3.5 Sonnet; if you’re optimizing for speed and throughput, consider Claude 3.5 Haiku.

Safety and usage tips: start with controlled scenarios

Anthropic notes that the new Claude 3.5 underwent strict safety evaluations before deployment, including collaborative testing with AI safety research organizations in the US and UK, and it still follows its ASL-2 standard. In real-world rollout, it’s recommended to start Claude 3.5 in “reversible” tasks—such as read-only browsing, draft data entry, or step recommendations—rather than directly handling high-risk actions like payments, destructive database operations, or permission changes. If you place approval checkpoints before critical steps, Claude 3.5’s computer use can save time without getting out of control.