Computer Use has been around since the 3.5 models — as a beta, with the charm of an experiment. Now it’s generally available on the Claude Platform, together with the Skills API and the Files API. And it comes with a new tool that tells me more about where things are heading than the GA checkmark itself.
Clicking like a human — but several actions per turn
The basic idea stays the same: you hand the agent a screenshot, it clicks, types, and scrolls like someone at the keyboard. The trick in the new version is that Claude now takes several actions per turn instead of one per model call. Fewer calls, less time, less cost. And Computer Use is now cleared for HIPAA-regulated workloads under the BAA — a signal aimed at healthcare and insurance.
The numbers Anthropic quotes are tangible. One customer in insurance reports that their longest claims workflow dropped from 32 to 13 minutes, cost per task fell about 30 percent, and completion hit 100 percent — with not a single prompt changed.
The browser tool no longer aims at pixels
The more interesting part is the new browser tool. It works with the same multi-action turns but also reads the structure of the page. Instead of guessing where a button sits on screen, the agent targets a specific field or link directly — through the page’s accessibility tree.
Anyone who has ever built a pixel-based agent knows the core problem: a new layout, a shifted button, and the whole script falls apart. Structure-based access makes that sturdier. Developers reach the tool via browser_toolset_20260801, and Computer Use runs as computer_toolset_20260801 — no beta header.
Skills and Files API in the same batch
The Skills API and the Files API are part of the same release, which I’ve touched on here before. Quick recap: skills are folders of instructions and scripts that Claude loads only when a task needs them — you host nothing yourself. The Files API stores the documents an agent reads and writes: upload once, reference by ID, download the finished files. Together the three pieces make an agent that operates software, applies your expertise, and hands back finished files.
My take
Of everything here, the browser tool is the most telling. Pixel-clicking was always the last resort for software without an API — impressive in the demo, brittle in daily use. Reading the page instead of staring at it is the right path, and interestingly Anthropic takes it right when others are still betting on plain screenshots.
That Computer Use is out of beta is more than cosmetics. It means Anthropic trusts the thing with productive workflows — including regulated areas like healthcare. The jump from 32 to 13 minutes is a single customer figure, sure. But the direction is right.
Sources: