OpenAI · 2025 · 01 · 23 · Tool · ~2 min read
OpenAI launched Operator (research preview)
What's actually new
- Browser automation any subscriber could try — the first time a big AI lab let normal people give an AI a task and watch it click around to do it.
- Cloud sandbox instead of your machine — different choice from Anthropic's Computer Use. Slower, but safer for new users.
- Takeover mode. If the agent gets stuck, you can step in mid-task without losing the context.
If you want more
Worth knowing
- Demo videos compress real time. Tasks took 5-15 minutes in practice, not the perceived 30 seconds.
- $200/month Pro tier was required at launch. Not Plus. High floor for casual testing.
- US-only at launch. Many people outside the US couldn't try it at all.
Who should care
Developers prototyping 'AI does things for you' apps. Operations teams considering AI for boring web tasks. Journalists and teachers tracking AI agent milestones. Anyone whose job involves a lot of repetitive browsing or data entry.
What to do about it
If you have Pro and live in the US, run one real task you'd normally pay an assistant for, and time it honestly. If you're outside the US, watch the demos critically — they hide the slow bits.
Honest take
Operator was the second big agent of the agent era, after Claude Computer Use. The cloud-sandbox design is more careful than Anthropic's 'control your real machine' approach — better for safety, worse for personal productivity. Both designs taught the same lesson: AI agents work, slowly and unevenly, on tasks where getting it wrong doesn't matter much. We're still 1-2 generations away from agents you'd trust with anything you can't undo.
Sources
- OpenAI — Introducing Operatorvendor
- Simon Willison — Operator, OpenAI's browser-using agentthird party
- Ben Thompson — The agentic momentthird party
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com