Anthropic has moved Computer Use, the Skills API, and the Files API to general availability all at once, and added a new browser tool inside Computer Use. All three had existed in beta, requiring developers to send a beta header and living under beta-tier availability and rate limits.
Computer Use still works the same way: the model gets a screenshot and clicks, types, and scrolls like someone sitting at a keyboard. The most practical change this time is multiple actions per turn. Previously, a single API call could only return one action, so even moving the mouse required a round trip. Now the model can issue several steps in one turn, cutting total task time accordingly. Computer Use is also now covered under Anthropic's BAA, meeting HIPAA requirements — clearing a compliance hurdle for healthcare and insurance use cases.
The Browser Tool Takes a Different Path
The new Browser Use tool extends computer control to web applications, but it works differently from screenshot-and-click: the model reads the page structure and locates specific elements directly, without relying on pixel coordinates.
That difference matters a lot more in production. A script that clicks by coordinates will hit the wrong spot the moment a window resizes, fonts render differently, or a cookie banner pops up — and it fails quietly. The model thinks it clicked “Submit,” but actually clicked blank space, and the whole workflow keeps running in a broken state. A model that locates elements gets information like “this is a button named submit,” which makes errors far easier to catch upstream.
Automation went through this exact transition once before, back in the Selenium era, moving from image matching to DOM selectors. This wave of agents started from screenshots and has now arrived at the same conclusion.
The two approaches aren't mutually exclusive, either. Screenshot-based control wins on coverage — desktop software, remote terminals, and legacy Windows clients have no DOM to read, so vision is the only option. The browser tool handles the web side, which happens to make up the bulk of enterprise back-office systems. Anthropic has folded both into the same Computer Use tool, so developers don't have to choose between two APIs — the model switches methods based on the type of window it's targeting.
Skills Move to an Upload Model
The Skills API addresses something different. Anthropic defines a skill as a folder containing instructions, scripts, and templates that the model loads only when a task needs it. The GA version simplifies uploading and version management, and skills now run inside Claude's own code sandbox, so developers don't need to set up an execution environment themselves.
Not having to host that environment is a bigger deal than it sounds. A team turning an internal report-generation workflow into a skill used to have to figure out which machine the script runs on, how to sandbox it, and how to clean up after failures. Now the platform absorbs that part. The trade-off is that the execution environment belongs to Anthropic — what dependencies can be installed and whether outbound network access is allowed follows the platform's rules.
File Storage Gets a 1TB Quota
The changes to the Files API read more like infrastructure tuning: files now support automatic expiration, the rate limit has been raised to five times its previous level, and each organization gets 1TB of storage. Once a file is uploaded, later requests can reference it by ID, so the same contract or financial report doesn't need to be stuffed into the context window again and again.
1TB is more than enough for text-heavy applications, but it's a different story for apps handling scanned documents, audio recordings, or video frames. The automatic-expiration part is easy to overlook, but it takes cleanup off developers' to-do lists — a long-running agent won't burn through its quota just because someone forgot to delete intermediate files.
Asteroid's Three Numbers
The customer example Anthropic cites is Asteroid, a company that automates insurance claims processing. Its longest claims workflow dropped from 32 minutes to 13, per-task cost fell by roughly 30%, and completion rate reached 100%.
Of the three numbers, completion rate is the hardest to move. A claims workflow crosses several back-office systems, and a single wrong click anywhere along the way can strand the whole process halfway through — at which point manual handoff costs more than not automating at all. The time and cost gains largely come from multiple actions per turn and fewer retries; the completion-rate improvement looks more like a result of the browser tool specifically.
Doing rough math on the published numbers: cutting a 32-minute workflow to 13 minutes roughly doubles how many claims the same team can process in a day, and a further 30% cost cut is what makes the per-claim economics work. Insurance claims have almost no tolerance for error — over- or under-paying a claim invites accountability — so a company willing to publish a 100% completion rate is effectively saying it lets the workflow run all the way through without a human gate.
All three features are currently available on the Claude Platform and Microsoft Foundry, with Google Cloud's Vertex AI to follow. Other platforms are catching up in the same direction around the same time — OpenAI previously merged Codex's reasoning and computer-use capabilities into a single model, a path that now intersects with this GA release. Whose control turns out more reliable will ultimately be decided by completion rates on workflows like this one that span multiple systems, not by benchmark scores.
Sources: Anthropic official announcement, CocoLoop, Claude Platform documentation; the rate-limit multiplier, storage quota, and the three Asteroid metrics were each checked individually.