One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

Claude Computer Use, Browser Tool, Skills API Reach General Availability

Four early-access features move to GA, round-trips cut by up to 40%

이미지: @ClaudeDevs (X) 영상 갈무리

Summary

  • Anthropic has moved Computer Use, the Browser Tool, the Skills API, and the Files API on the Claude Platform to general availability (GA)
  • The new Computer Use toolset handles multiple actions per turn, cutting round-trips by 20-40% for early-access customers
  • Asteroid's healthcare automation tests showed a 32-52% drop in model calls, 25-32% lower costs, and a 100% completion rate
Video from the source
발표
컴퓨터 사용·브라우저 툴·스킬 API·파일 API, 클로드 플랫폼 정식 출시(GA), 2026-08-20
컴퓨터 사용 툴셋
computer_toolset_20260801 — 턴당 여러 동작 처리로 라운드트립 20~40% 감소(얼리 액세스 기준)
브라우저 툴셋
browser_toolset_20260801 — 좌표+요소 참조 정보로 레이아웃 변경에도 동작
파일 API 업데이트
expires_in_seconds 파라미터 추가, 요청 한도 분당 500회(5배), 조직당 저장 용량 1TB
스킬 API
절차를 업로드해 버전 관리, version_id 또는 latest로 요청 고정
의료 자동화 사례
애스터로이드 — 모델 호출 32~52% 감소, 비용 25~32% 절감, 워크플로 완료율 100%

Multiple actions in a single request

Previously, whenever Claude operated a screen, every click and every keystroke required a new request to the server. With the new computer_toolset_20260801 version, clicks, text input, key presses, and screenshots can now be bundled into a single turn. Anthropic's developer account ClaudeDevs said this change cut per-task round-trips by 20-40% for early-access customers. Fewer round trips mean tasks finish faster and cost less.

With this update, Anthropic has moved four features on the Claude Platform — Computer Use, the Browser Tool, the Skills API, and the Files API — to general availability (GA). This means capabilities that had been limited to early access are now available without a separate application process. The Claude Platform has also been the site of a steady stream of additions in recent weeks, including budget caps, an advisor model, and memory and domain controls.

FeatureNew version/changeKey effect
Computer Usecomputer_toolset_20260801Multiple actions per turn, round-trips ↓20-40%
Browser Toolbrowser_toolset_20260801Element references + coordinates handle layout changes
Skills APIVersion managementFixed procedures via version_id/latest
Files APIexpires_in_seconds, 500RPM, 1TB/orgExpanded limits for file reuse
이미지: @ClaudeDevs (X)

Reading screen structure instead of pixels

A chronic problem with browser automation has been that even a minor change in screen layout could break a script. Because clicking relies solely on coordinates, if a button moves, the automation ends up clicking the wrong spot. The new browser_toolset_20260801 now provides element reference information within the page in addition to coordinates. According to the Computer Use tool documentation, Claude can use this information to find and click the same button even when the layout changes. This approach is especially useful for automating internal systems that lack APIs — that is, in-house programs with no external integration point.

Saving procedures and managing versions

The Skills API and Files API are building blocks for creating Claude Managed Agents. The Skills API lets a team upload a recurring work procedure once, assign it a version, and then reference either a specific version_id or the latest version (latest) in subsequent requests. The Agent Skills documentation explains that even after a procedure changes, previous versions can still be referenced. The Files API lets a file uploaded once be reused across multiple requests via a file_id; this update adds an expires_in_seconds parameter to set expiration times, raises the per-minute request limit fivefold to 500, and increases per-organization storage capacity to 1 terabyte.

Numbers from healthcare automation in the field

David Mlcoch of Asteroid said his company applied the new Computer Use tool early to a healthcare business automation workflow through early access. According to figures he shared, model calls dropped 32-52%, per-task costs fell 25-32%, and every workflow tested achieved a 100% completion rate. Anthropic also published a post on the Claude blog explaining how to combine these four features to build production-grade agents.

How to try it

① Where to start — Get started by issuing an API key from the Claude Platform developer console (platform.claude.com). All four features are used via API calls, so there are no separate buttons in an app interface.

② Step-by-step usage

  1. Issue an API key and include it in the request header.
  2. For tasks involving screen operations, specify computer_toolset_20260801 or browser_toolset_20260801 as the tool in your request.
  3. For recurring work procedures, upload them once via the Skills API, then reference the relevant version_id or latest in subsequent requests.
  4. For frequently used files, upload them once via the Files API to obtain a file_id, and if needed, set an expiration with expires_in_seconds for reuse across multiple requests.

③ Who can use it — This is aimed at developers and companies with access to the Claude Platform API. Now that these features are GA, they can be used simply by specifying the toolset names given in the documentation, without the separate early-access approval process required before.

④ What you can try — Use cases include automating recurring tasks in programs without APIs, such as internal systems, or attaching automation to standardized document-processing workflows to speed things up. The Skills API's version-locking feature is also useful when multiple teams need to run different versions of the same procedure simultaneously.

Editor's view

This announcement feels like the capstone to weeks of updates. From budget caps and an advisor model on August 9, to memory and domain controls on August 19, to AG-UI integration on August 20, Anthropic has been adding a new piece to Claude Managed Agents almost every week. The GA release of Computer Use, the Browser Tool, and the Skills/Files APIs signals that these pieces have shed their early-access label and are now ready for production use.

Anyone who has worked with screen automation knows how often pixel-coordinate-based automation breaks. If a button shifts by even a few pixels, a script clicks the wrong spot, requiring the coordinates to be reset each time. The approach of providing element reference information alongside coordinates substantially reduces this repetitive maintenance burden. The reported 20-40% cut in round-trips is best understood, ultimately, as meaning the same screen no longer has to be recaptured over and over.

For teams in Korea, the point worth noting is internal systems without APIs. Many organizations still run legacy ERP or groupware systems that offer no external integration point, and automating these has historically meant relying on visual screen operation. The Asteroid healthcare automation case — with model calls cut nearly in half and costs down by more than a quarter — offers a useful benchmark for organizations with standardized document-processing work. That said, the success of version-managing procedures through the Skills API will ultimately hinge on how thoroughly those procedures were documented from the start.

In the coming weeks, more examples of managed agents combining these four features are likely to emerge. The pieces are now in place as official features: attaching screens via AG-UI, locking procedures with the Skills API, and reusing documents through the Files API.

Comments