Anthropic's Computer Use feature, which lets Claude autonomously operate a desktop browser the way a human would, has exited public beta. The launch coincides with announcements from Stripe and Shopify that they have moved internal automation workflows from custom RPA scripts to Claude-driven agents in production.
The shift represents the first real-world enterprise deployment of an agentic AI feature at meaningful scale. Stripe is using Computer Use to automate parts of its merchant onboarding review process — a workflow that previously required human operators to navigate four separate internal tools. Shopify, meanwhile, has deployed it across customer support, where Claude now handles roughly 18% of tier-one ticket volume end-to-end.
How it actually works in production
The version exiting beta is materially different from the demo Anthropic showed in late 2024. The model now operates with what the company calls 'verifiable action chains' — every click, keystroke, and form submission is logged with screenshots, allowing audit trails that satisfy SOC 2 and PCI compliance reviewers. This was the single biggest blocker preventing enterprise adoption in the previous version.
Reliability has also improved sharply. Anthropic says the latest version completes multi-step browser tasks at a 94% success rate on its internal benchmarks, up from 73% at the original beta launch. Customer-reported numbers track closely: Stripe's internal evaluations put the production success rate at 91% for tasks of moderate complexity, with the failures concentrated in legacy interfaces that lack proper accessibility tagging.
What it means for the agent market
The launch puts pressure on competitors who have been demonstrating similar capabilities but have yet to ship them as production-grade products. OpenAI's Operator, while available to consumers, is still positioned as an exploratory product rather than enterprise infrastructure. Google has shown demos of Project Mariner but has not announced a general-availability date. For now, Anthropic appears to have a quarter or two of head-room before the field catches up — a meaningful but probably not decisive lead.
