Anthropic's Browser Use avoids pixel-based navigation with accessibility references

Anthropic has released Browser Use, a tool that enables Claude to navigate and interact with web pages in a fundamentally different way than its earlier Computer Use capability. Rather than analyzing page screenshots to determine button positions and calculate click coordinates, Browser Use gives Claude direct references to page elements.

How it differs from Computer Use#

Computer Use operates by examining desktop screenshots and translating them into mouse and keyboard commands using pixel coordinates. Browser Use operates within the browser itself, leveraging the page's underlying structure rather than trying to reverse-engineer element locations from visual data alone.

When Claude calls read_page, the system returns a text representation of the accessibility tree—a structured view of links, buttons, input fields, and other interactive elements. Each element receives a reference identifier (such as ref_3). When Claude wants to click that button, it simply passes the reference ID rather than calculating screen coordinates.

This approach avoids the fragility of coordinate-based interaction, though it does introduce a new consideration: if the page navigates or updates significantly, a reference that was valid moments ago may become stale. The developer's executor must detect such mismatches, reject the outdated action, and prompt Claude to re-read the page before continuing.

Batching reduces model calls and costs#

A major efficiency gain comes from allowing multiple browser actions in a single model request. Instead of Claude making a request, waiting for a response, then making another request for the next action, it can now bundle several operations together—click a button, fill a field, submit a form—and send them all at once.

The executor processes these actions sequentially, since each step depends on the outcome of the previous one. This batching cuts model calls significantly, which matters as workflows grow from a handful of interactions to dozens or hundreds. Reducing back-and-forth exchanges lowers both latency and cost, particularly important given that longer browser tasks can quickly accumulate expensive model invocations.

Developer-hosted sessions, not Anthropic-hosted#

Unlike some other tools in Anthropic's recent release, Browser Use does not run inside Anthropic's infrastructure. The developer must provide and maintain their own browser environment, handle session persistence between turns, and translate Claude's requests into actual browser operations. The tool adds approximately 6,600 input tokens to each request before accounting for accessibility tree data and screenshots sent back to the model.