For most of its life, ChatGPT was one box you typed into. That is no longer true. OpenAI has reorganized ChatGPT into three distinct surfaces, each tuned for a different kind of job, and the shift changes how you decide where to take a task. Understanding the new ChatGPT means understanding ChatGPT Chat, Work, and Codex, the three modes that now sit under one name.
The short version is easy to remember: Chat is for answers, Work is for workflows, and Codex is for development. They share the same underlying model family, but they are built around different assumptions about what you want back, how long the task runs, and how much of it the software should carry on its own. This piece walks through what each surface is, where the browser went in the process, and how to choose between them without overthinking it.
Why OpenAI split ChatGPT into three
A single chat window is a wonderful thing for a quick question and an awkward thing for a job that takes twenty steps across five apps. As OpenAI pushed ChatGPT from a place you ask things into a place that does things, the one-window design started to strain. Answering a question, running a multi-step workflow, and editing a codebase are genuinely different activities with different needs around context, permissions, and how long the software should keep going before it checks back in.
Rather than bolt every capability onto the same interface, OpenAI separated them into three products with clear jobs. The model underneath is the same modern family (the GPT-5.6 generation, which OpenAI positions as its strongest for coding and agentic work), so the intelligence does not change when you move between surfaces. What changes is the scaffolding around it: how much reach it has into your apps and files, how autonomously it runs, and what a finished result looks like.
ChatGPT Chat: the fast path for answers
Chat is the ChatGPT you already know, and it is the fastest path when you want an answer rather than a finished deliverable. It is built for the quick loop: ask a question, brainstorm an idea, draft or rewrite a paragraph, work through a problem out loud, get unstuck. You stay in the conversation, you steer as you go, and the value shows up in seconds.
Chat is the right default for the large share of everyday tasks that do not need to touch your connected systems or run in the background. If the job is "help me think, write, or understand something right now," Chat is where it belongs. It is also where features like company knowledge in ChatGPT and the real-time voice mode in ChatGPT Live live, so the conversational surface has grown well beyond plain text without changing its core promise of a fast answer.
ChatGPT Work: for multi-step workflows
Work is the surface for jobs that span several steps and produce a finished artifact. Instead of returning instructions for you to follow, it takes a goal and turns it into completed materials: documents, spreadsheets, reports, and similar outputs. It is available on web, mobile, and desktop, and it is powered by the same agent runtime that drives OpenAI’s coding agent, which is why it can operate across your connected apps and files rather than staying trapped in a single chat.
The distinction that matters is autonomy over time. Work can carry a task across multiple steps, pull the context it needs from the systems you have connected, and hand back a result rather than a recipe. We cover this surface in depth in our explainer on ChatGPT Work, and because it competes directly with Anthropic’s equivalent, our piece on ChatGPT Work versus Claude Cowork is the better read if you are deciding between the two agentic-work tools. The rule of thumb for Work is simple: reach for it when the finish line is a deliverable, not a reply.
Codex: for repo-level development
Codex is the surface for building software. It works at the level of a real repository, with access to local code, terminals, and developer tooling on the desktop, and it now runs on mobile as well so long-running work can move with you. It is not a chat window that happens to know how to code. It is a development environment where the agent can read the repo, make changes, run commands, and keep a task going until the checks pass.
Codex has also absorbed a growing set of developer capabilities. It can edit diffs and review pull requests, with follow-up edits handled in place. It ships hardware companions and mobile clients too, including OpenAI’s first physical accessory, the macropad we covered in the Codex Micro, and the phone client in Codex on mobile. If your task involves a codebase, tests, or a terminal, Codex is the surface designed for it, and the other two will feel like the wrong tool.
Where the browser went
If you are wondering what happened to OpenAI’s browser in all of this, the answer fits the same pattern. OpenAI retired its standalone browser and folded browsing into these surfaces instead of keeping it as a separate destination. We told that story in OpenAI retiring the ChatGPT Atlas browser: the company decided a whole browser was a feature rather than a product, and split the capability across a browser built into the desktop app, a Chrome extension, and a cloud browser that runs remotely.
Codex is where the most capable version of that browsing now lives. It carries a built-in browser that works across signed-in tabs, downloads, and annotations, a Chrome extension for using your existing browser session, and a computer-use ability for longer multi-step tasks. In other words, the browser did not disappear. It became a feature inside the surfaces, which is exactly the direction we described in our look at the agentic browser landscape, where the browser increasingly acts rather than just displays.
How ChatGPT Chat, Work, and Codex differ
The three surfaces are easiest to keep straight by the shape of what you want back. Chat returns an answer inside a conversation. Work returns a finished deliverable produced over several autonomous steps. Codex returns changes to a codebase, made inside a real development environment. The deeper you go down that list, the more reach and autonomy you are handing the software, and the longer a single task tends to run before it needs you again.
There is honest overlap at the edges. Chat can draft a short document, and Work can answer a question on the way to building something. The point of the split is not a hard wall between them but a center of gravity for each: a natural home where a given kind of task runs best. Because the same GPT-5.6 model family sits under all three, moving a task from one surface to another does not cost you intelligence. If you want help matching the model tier to the job as well, our practical guide to GPT-5.6 covers the Sol, Terra, and Luna options.
What the split signals
Read together, the three surfaces describe where OpenAI thinks its product is heading. The company is betting that people will use ChatGPT in three fundamentally different modes, and that trying to serve all three from one interface holds each of them back. Splitting them lets each surface make sharper assumptions: Chat can stay fast and light, Work can lean into autonomy and connected apps, and Codex can behave like a full development environment rather than a chat with a code block.
It is worth keeping OpenAI’s framing at arm’s length. The company describes its latest models as setting a new standard and, by its own account, outperforming rival models while using fewer tokens. Those are OpenAI’s claims about its own products, and the honest way to read them is as marketing until independent testing weighs in. What is not in dispute is the structural change: ChatGPT is no longer one box. It is three doors, and knowing which one to open is now part of using it well.
Frequently asked questions
What are the three parts of the new ChatGPT?
The new ChatGPT is organized into three surfaces: Chat, Work, and Codex. Chat is the conversational surface for fast answers, brainstorming, and writing. Work is an agent that runs multi-step workflows across connected apps and files and returns a finished deliverable. Codex is a development environment for repo-level coding with local code, terminals, and developer tools.
What is the difference between ChatGPT Chat and ChatGPT Work?
Chat returns an answer inside a conversation and is best for quick, everyday tasks that do not need to touch your connected systems. Work takes a goal and produces a finished artifact, such as a document, spreadsheet, or report, by running several steps on its own across the apps and files you have connected. Use Chat when you want a reply, and Work when you want a deliverable.
When should I use Codex instead of Work?
Use Codex when the task involves a codebase, tests, or a terminal. Codex operates at the level of a real repository, with access to local code and developer tooling, and it can make changes, run commands, and review pull requests. Work is aimed at general knowledge-work deliverables rather than repo-level software development.
Do Chat, Work, and Codex use the same model?
Yes. All three surfaces run on the same modern model family, the GPT-5.6 generation. Moving a task between them does not change the underlying intelligence. What changes is the scaffolding around the model: how much access it has to your apps and files, how autonomously it runs, and what a finished result looks like.
What happened to OpenAI’s browser?
OpenAI retired its standalone ChatGPT Atlas browser and folded browsing into its other surfaces. The capability now lives as a browser built into the desktop app, a Chrome extension, and a cloud browser, with the most capable version inside Codex, which carries a built-in browser and a computer-use ability. The browser became a feature inside the products rather than a separate destination.
Is Codex available on mobile?
Yes. Codex runs on the desktop with local code, terminals, and developer tools, and it also has a mobile client so longer-running development work can continue from a phone. OpenAI has also released hardware for Codex, including a macropad accessory designed to control the coding agent.
Which surface should I use by default?
Match the surface to the shape of the result you want. For a fast answer or a bit of writing and thinking, use Chat. For a multi-step job that should end in a finished document, spreadsheet, or report, use Work. For anything involving a repository, use Codex. There is some overlap at the edges, but each surface has a natural home where its kind of task runs best.
