Their ‘Computer Use’ model took over a real laptop, ordered delivery, and handled payment - autonomously.
Anthropic didn’t just ship another chat model - they built an AI that controls your entire computer like a human. Cursor movement, clicking buttons, typing passwords, even scrolling PDFs. Their demo? Agent logs into DoorDash, picks pizza, pays with Apple Pay, all unsupervised.
This isn’t cute - it’s terrifyingly practical. Forget API wrappers; this is AI using your EXISTING apps. As a dev, imagine deploying agents that onboard themselves to GitHub, merge PRs, deploy to Vercel, or debug prod issues by SSHing into servers. The tooling explosion we predicted? Here.
Honest opinion: Claude’s agentic leap lapdogs GPT’s o1-preview. But scaling this safely? Massive unsolved problem. One rogue click and your AWS bill hits $10k. Still, for controlled environments (CI/CD, testing), this rewrites automation.
Would you let this loose on your prod machine?
Source: Anthropic