Anthropic puts Claude's computer use into full production

If you have ever tried building an AI agent to navigate an internal portal without an API, you know the pain: it takes a screenshot, guesses a coordinate, clicks, waits for another model call, and often misses the target completely.

Anthropic is rolling out three production-ready tools to make this workflow actually reliable. Computer use, the Skills API, and the Files API are now generally available on the Claude Platform, alongside a dedicated browser use tool.

Precision web navigation and multi-action turns

The original computer use feature worked purely on vision, treating the monitor like an image canvas. The new browser use tool pairs those screenshots with actual page structure, targeting specific fields and buttons rather than blind coordinates.

Claude can also now execute multi-action turns. Instead of requiring a round-trip model call for every single click, it strings together several actions in a single turn.

The performance gains in production are hard to ignore. In complex insurance workflows lacking APIs, Anthropic reported processing times dropping from 32 minutes to 13 minutes. Task costs fell roughly 30% with a 100% completion rate, all without prompt modifications.

Anthropic also added HIPAA eligibility for computer use under its business associate agreement, unlocking healthcare data workflows.

Packaged team skills and 1 TB file handling

The other two APIs handle institutional memory and heavy documents.

The Skills API lets developers upload versioned folders containing instructions, scripts, and templates. When Claude encounters a relevant task, it pulls the skill and executes it inside Anthropic's secure code sandbox, removing the need to host external execution runtimes.

Box is already using this within Box Agent to package banking credit methodologies, letting the agent analyze financial statements and draft memos directly.

Meanwhile, the Files API gives every organization 1 TB of storage with 5x higher rate limits and automatic expiration. Teams upload spreadsheets or PDFs once and reference them by ID, rather than re-uploading massive files on every prompt.

Both the Skills API and Files API are also live on Microsoft Foundry, with the computer and browser tools coming soon to Google Cloud Vertex AI.

Why it matters

We think Anthropic is making the right bet by prioritizing enterprise infrastructure over flashy benchmark claims.

For developers, this dramatically changes the economics of legacy automation. Cutting task times by more than half while reducing costs by 30% makes automating tedious web portals financially viable today, especially with the added safety of DOM-based element targeting.

Packaging sandboxed code execution, versioned company playbooks, and large-file storage directly into the platform saves teams weeks of custom plumbing. You no longer need to manage custom orchestration just to let an agent inspect a spreadsheet and click a web form.

The lingering challenge will be dynamic interfaces. While structural targeting beats raw pixels, complex single-page apps and iframe-heavy enterprise software can still trip up DOM parsers, meaning human oversight remains necessary for high-stakes workflows.

Still, this sets a high bar for agent tooling. Anthropic is systematically turning Claude from a conversational model into an end-to-end digital worker.