ChatGPT Desktop Becomes a Workspace: What Chat, Work and Codex Change

The era of typing a question into a single box and waiting is over. OpenAI has turned the ChatGPT desktop app into a three-lane workspace: Chat for quick questions, Work for deliverables that take hours, Codex for software tasks. The narrator walks through the whole change with one clear thesis: you no longer ask questions, you delegate entire jobs.

The logic of the split is simple and strong. Chat exists for brainstorming, quick questions and instant help; the familiar talk experience stays exactly where it was. Work is a goal-locked agent: it researches a topic, analyzes information and produces finished pieces such as documents, sheets, decks, reports and sites. According to OpenAI Help Center docs, Work can split a goal into steps and run on its own for hours when needed; you step in to answer questions, change direction and approve important moves.

Projects and context: your notes now live on desktop

Projects arrive on desktop with the same arrangement you already use on the web. Notes, files, older pieces and case notes gather in one context; from the same project you ask Chat for a draft in seconds and Work for a full piece built on project context. In the narrator's example, video outlines come first from community notes, then a landing page from the same context. For projects bound to a local folder, the agent can also reach files on your machine to the extent you permit.

The continuity layer is built from two parts. The Recents view gathers Chat and Work chats in one list; you filter, pin and find old jobs in seconds. The cross-device rule is crisp: Cloud Work chats sync across web, mobile and desktop, so you start at the desk and continue on the phone. According to TheWinCentral, this synced history plus sidebar projects became the most loved part of the update. Local chats stay on that computer; knowing which mode you are in matters for privacy.

Do not skip the rollout note: Work opens gradually and not every account gets access on the same day. OpenAI keeps the old app under the ChatGPT Classic name; the classic build keeps receiving model updates and security patches while new agent powers land in the new app first. On the enterprise side, admins can switch Work access per role. So if the mode switcher is missing after your update, do not panic, retry in a few days and confirm access with your admin.

Speed layer: keyboard shortcuts and the screen

Daily speed comes from keyboard shortcuts. On the Mac, Option + Space opens a small chat box from any screen; you ask for headline ideas in the middle of an email, an edit or a page draft and return to the flow. On Windows, stay careful: the key sequence described in the talk does not match the Alt + Space note inside OpenAI Help Center one to one, so confirm the shortcut field in Settings on your own install. The September 2026 notes inside ChatGPT Learn also confirm that quick questions can start from New chat in the Codex view.

The Windows-only Appshots feature runs on a double Alt press: it takes a screenshot of the app you use and carries it straight into the chat. The agent, seeing what you see, reads the layout and gives advice. It strengthens visual jobs such as ads, landing pages or thumbnails; instead of describing what feels off, you show it. A small but oft-used bridge that removes the chore of turning screen context into words.

The built-in browser lets you research without leaving the app: Command Shift B on the Mac, Control Shift B on Windows. You pull a tab into a Work or Codex chat and keep working where you are already signed in; comparing a rival page against your project notes and asking what you miss is the classic example. According to ChatGPT Learn docs, browsers such as Edge, Brave, Opera and Vivaldi are supported too, and site actions can plug into the agent through WebMCP site tools.

Voice and code: steer by talking, ship in a separate lane

Voice is the overlooked piece. It comes built into Work and Codex : you start a request by talking and cut in mid-run to fix the direction. OpenAI Help Center describes corrections such as a no-coding-tools filter landing naturally mid-task; this is a true agent dialogue, not dictation. Technical note: only one voice conversation runs at a time, and you must grant microphone plus screen-recording permissions up front.

The last lane belongs to developers. Codex sits in a separate view working with local folders, repositories, terminals and developer tools; its history stays apart from Chat and Work history. According to the Codewalkers guide, the desktop view is the easiest door for anyone who hands it a folder to start an agent task with no terminal memory. The open-source work on GitHub describes inline diff comments and review, pull-request review in the side panel and an approve flow on clean branches. Those who want scale can run parallel jobs on worktrees without collisions.