OpenAI Codex hits 90M installs

Developer AI just took a step closer to the real workplace browser. That’s the actual news here. OpenAI did ship a Codex extension for Chrome this week, and it lets the agent work inside the tabs where people already do sales ops, support work, QA, and web testing. But the viral “90 million installs” angle appears to be wrong. ### What actually launched? OpenAI added a Chrome extension for Codex on May 7, 2026. The extension gives Codex access to browser tasks that need your signed-in session — things like Gmail, Salesforce, LinkedIn, dashboards, or internal tools. The point is simple: a lot of real work lives behind logins and inside tabs, not inside clean APIs. ### Why does Chrome matter so much? Because browser state is where the messy stuff is. A coding agent can already edit files, run tests, and use APIs. But plenty of useful work still depends on cookies, account permissions, page structure, and whatever is open in your browser right now. OpenAI says Codex can now research, fill forms, review dashboards, and move through multi-step workflows in task-specific tab groups without taking over your active session. (developers.openai.com) ### So did Codex really hit 90M installs? No public evidence supports that. The Chrome Web Store listing for Codex showed about 80,000 users when checked today, with 97 ratings and version 1.1.4 updated on May 7, 2026. If there is some larger internal usage metric floating around, OpenAI has not publicly tied it to Chrome installs. The public number is five digits, not eight. (chromewebstore.google.com) ### Where did the confusion likely come from? Turns out there are several different “Codex” surfaces now — the Codex app, CLI, IDE extension, cloud agent, GitHub action, and now Chrome. It’s easy to mash those together into one giant adoption number. OpenAI has talked before about fast Codex growth and heavy internal use, but that is not the same thing as one browser extension suddenly reaching mass-market scale. (chromewebstore.google.com) ### What is Auto-review? Auto-review is the more interesting second half of the story. OpenAI released it in late April as a middle mode between constant human approvals and full unrestricted access. Instead of stopping every time Codex wants to cross a boundary, a separate reviewer agent checks whether the action looks safe enough to proceed. ### Why is that a big deal? (openai.com) Because approval friction kills long-running agent workflows. OpenAI says Auto-review makes Codex stop for human approval roughly 200x less often than manual mode. In one internal snapshot, 720 actions that would have interrupted a user were reviewed automatically, 713 were approved, and only 7 were denied. That is the difference between “useful demo” and “thing people actually leave running.” (alignment.openai.com) ### What’s the catch? The catch is access. OpenAI’s docs are pretty explicit that browser content is untrusted, and that history, internal URLs, search terms, authentication data, and page content can all be sensitive. Users can allow one site for one chat, always allow a host, or go further and enable elevated-risk settings. Basically, the product is more useful precisely because it gets closer to the dangerous parts of your workflow. (alignment.openai.com) ### Bottom line This story matters, but not because Codex somehow got 90 million installs overnight. It matters because OpenAI moved its agent from the sandbox toward the logged-in browser — and paired that jump in power with a system designed to reduce approval drag without going fully hands-off. That is a real shift in how AI tools start becoming everyday work software. (developers.openai.com)