OpenAI just turned ChatGPT from a chatbot into something that drives your computer while you watch — and the training data problem that held this back for years quietly got solved.
The Summary
- OpenAI is rolling out Computer Use tools across ChatGPT, including a Chrome extension, cloud browsers, and integration with apps beyond just coding environments
- The breakthrough: OpenAI finally collected enough training data to teach agents how to navigate computers and browsers with human-level dexterity, a problem that's blocked this since 2015
- Over 600,000 organizations already testing rival Anthropic's Claude Cowork feature as of May — OpenAI is playing catchup but betting ChatGPT's scale wins
- President Greg Brockman says April update made the tech work "for anyone who does computer work," not just coders
The Signal
The race to build agents that can actually use your computer just shifted from research problem to deployment sprint. OpenAI spent eleven years stuck on the same bottleneck: you can't teach an AI to navigate software interfaces without massive amounts of training data showing how humans actually do it. Click here, scroll there, fill this form, wait for that page to load. That dataset didn't exist at scale until recently.
Now it does. OpenAI won't say exactly how they built it, but the rollout timeline tells the story. Chrome extension in early 2026. Cloud browsers for public website interaction. Computer Use added to Codex. Then the expansion beyond coding tools to general app control. That's not a research lab tiptoeing into beta — that's a company with enough training data to ship production tools to millions of users.
"Once ChatGPT can use computers and software faster than you or I can, it's going to change the way that you, by default, want to interact with your computer." — Ari Weinstein, OpenAI
The speed advantage is the entire value proposition here. An agent that controls your browser and apps doesn't need to be smarter than you. It needs to be faster and tireless. If it can execute a 47-step workflow in three minutes while you're in a meeting, it doesn't matter that you could have done it better manually in forty minutes. You're never doing it manually again.
Anthropic shipped first in 2024 and already has 600,000 organizations testing Claude Cowork. That's a meaningful head start. But OpenAI has distribution that Anthropic can't match — ChatGPT's existing user base dwarfs Claude's. The question isn't whether people will adopt computer-use agents. It's whether Anthropic's technical lead or OpenAI's scale advantage wins the default position.
The tells that this is real:
- OpenAI staff calling it an "inflection point" instead of "early research"
- Integration across multiple surfaces (browser, cloud, apps, coding tools) instead of one narrow demo
- Language shifting from "AI assistant" to "automate computer work"
The Implication
Watch what gets automated first. It won't be creative knowledge work. It'll be the repetitive multi-app workflows that eat 30% of your week — data entry across systems, report generation, scheduling chains, CRM updates. If you do computer work that follows similar patterns week to week, your job description just got a rewrite.
For companies: the build vs. buy decision on internal automation tools just tilted hard toward buy. Why build custom RPA workflows when ChatGPT can watch you do a task once and replicate it?