Anthropic ships computer use, browser tool, Skills API to GA

Anthropic pushed four Claude Platform components out of preview on August 20: computer use, a browser tool, the Skills API, and the Files API. The company’s developer account on X announced the general availability release at 10

PM that evening, framing the update around agents that operate software lacking a formal API.

The computer use change is architectural. Claude can now issue several actions per turn (clicking, typing, pressing keys, taking screenshots) instead of waiting for a round trip after each one, under a tool version Anthropic labels computer_toolset_20260801. Round trips per task dropped 20 to 40 percent among Anthropic’s early-access customers, the company says, letting those workflows complete quicker and at a lower cost. That figure comes from Anthropic’s own early-access program, not an independent benchmark, and the company has not disclosed the number of customers or tasks it covers.

The browser tool addresses a failure mode anyone who has run browser automation at scale already knows well: scripts that click on fixed pixel coordinates snap the moment a page redesigns a button or reflows a layout. Under the new browser_toolset_20260801 identifier, Claude receives the page’s underlying structure and can target actual page elements rather than guessing where they sit on screen. For a team maintaining these agents, that shifts the failure mode from silent breakage after every front-end deploy to something closer to normal software maintenance: an element reference that stops resolving throws an error you can catch, instead of a click landing on the wrong button with no signal that anything went wrong.

Anthropic positions the Skills API and Files API as the infrastructure layer under what it calls Claude Managed Agents. The Skills API lets a team upload a procedure once, assign it a version, and have each request pin to a specific version_id or float to latest. That is a meaningful operational detail: it means a team can update a skill’s instructions without touching every workflow that depends on it, or lock a production agent to a known-good version while testing a new one elsewhere.

The Files API works the same way for data. A file gets uploaded once and referenced by a file_id across as many subsequent requests as needed, rather than re-uploaded each time. The general-availability release adds an expires_in_seconds parameter for setting file lifetimes, raises rate limits fivefold to 500 requests per minute, and sets a 1 TB storage ceiling per organization.

None of this comes with independent verification. The announcement is Anthropic marketing its own product on its own developer account, and the 20 to 40 percent figure, the 1.4 million views the post drew, and every other number in the thread are the vendor’s claims about the vendor’s product. The linked blog post that would normally carry more detail was not available at the time of this report.

What changes for builders is less about any single number and more about the shift from experimental to production posture. Versioned skills and reusable files are the kind of primitives teams ask for once they are running agents that touch real customer workflows rather than demos, and Anthropic pairing this with GA status for computer use and the browser tool suggests it wants those agents to be trusted with longer, less-supervised sessions. Teams currently running computer-use pilots on the older toolset should benchmark the new element-reference approach in their own environment before the next production release, since Anthropic’s round-trip savings figure reflects its own customer sample, not a general guarantee.

Anthropic’s developer account on X (ClaudeDevs) posted this announcement on August 20, 2026.