OpenAI Launches ChatGPT Work and Folds Codex Into One Desktop App
ChatGPT Work is an agent that operates across your apps and files for hours to deliver finished documents, spreadsheets, and web apps. OpenAI also merged Codex into a unified desktop app and began sunsetting the Atlas browser.
OpenAI launched ChatGPT Work on July 9 — an agent that takes action across your apps and files, stays on a project for hours, and turns a goal into finished output: spreadsheets, slides, documents, even working web apps. It ships alongside a redesigned ChatGPT desktop app for Mac and Windows that absorbs Codex as a built-in mode, and it marks the beginning of the end for Atlas, OpenAI’s browser, which the company is now sunsetting.
The pitch is blunt: stop chatting with AI, start delegating to it. Where ChatGPT answers questions, Work breaks a complex project into smaller steps and completes them independently, checking in only when it needs a decision. It connects to roughly 15 third-party integrations at launch — Google Drive, Slack, Salesforce, and GitHub among them — addressable through an @-mention workflow. An Auto-Review feature scans agent actions for risky operations before they execute.
The engine underneath is GPT-5.6, which is rolling out at the same time. Work is available now for Pro, Enterprise, and Edu subscribers on mobile and web, with Plus and Business users following over the next few days.
The desktop consolidation matters as much as the agent. Codex — which OpenAI says now has more than 5 million weekly users, over 1 million of them doing non-software work — stops being a separate app and becomes one of three modes (Chat, Work, Codex) in a single ChatGPT desktop client, available globally on every plan including Free. One app, one login, three levels of delegation.
The timing was not subtle. Anthropic shipped Claude Cowork for mobile and web the same day, freeing its own agentic sessions from the desktop. Both companies converged on the identical thesis in the identical week: the next battleground is not chat quality but finished knowledge work — who can take a vague goal at 9 a.m. and hand back a completed deliverable by lunch.
For developers, the interesting detail is what happened to Codex. Merging it into the main app rather than keeping it as a specialist tool signals that OpenAI sees coding agents as a feature of general work, not a separate product category. The 1 million non-developers already using Codex for other tasks apparently made the decision for them.
The open question is billing. Agentic sessions that run for hours consume serious compute, and OpenAI is attaching usage-based pricing to heavy Work sessions on top of subscriptions. Whether customers accept metered agents as the new normal will shape how every lab prices this category.