ChatGPT modes: Chat, Work or Codex?

Chat, Work, and Codex are not three intelligence levels. They are three surfaces with different tools, different histories, and different ideas of what “done” looks like. Choosing the wrong one looks like a bad prompt: the model seems to ignore your file, refuse to touch your code, or write a report when you wanted a sentence.

If you typed chatpt looking for the product name, that is a separate question: chatpt is a typo, not a mode. This page is about the three places you can work inside ChatGPT.

The descriptions below were checked against OpenAI’s own “ChatGPT Work and Codex” help article on 31 October 2026. Mode names and availability change; the official page is the source, and this page dates its claims for that reason.

The one rule

Pick by the artifact and the surface, not by “which one is smartest”.

• If the output is a thought you want to develop in conversation, use Chat.
• If the output is a finished file someone will review, use Work.
• If the output is a change to a repository, use Codex.
• If you need to talk while you work, use Voice on whichever surface supports it.

The mode name describes the room, not a rank. A bad brief in Work is still a bad brief. A good brief in Chat can still be enough for a quick decision.

What each one is for

OpenAI’s own framing is unusually clear:

Work can also run scheduled tasks, monitor for changes, and pause for your approval. That is not a malfunction; it is the agent model. Codex is the one with the repository and the developer views.

Where each one lives

This is where most “it is not working” confusion comes from.

• Chat is available in ChatGPT on web, mobile, and the desktop app.
• Work is available in ChatGPT on web and mobile for eligible paid plans, and in the desktop app when your plan and workspace include it. Cloud Work chats sync across surfaces; local chats stay on the computer.
• Codex is a separate view in the desktop app. Codex Cloud can start or continue coding tasks from desktop, web, or mobile, but Codex itself is not selectable on web or mobile. The mobile app’s Remote tab can open supported desktop Codex chats; those do not become web or mobile chat history.
• Voice is in Work on web and mobile, and in Work and Codex in the desktop app on macOS and Windows.

Availability depends on the plan, region, rollout, and workspace settings. Do not assume that a mode you see on one surface is available on another.

The symptom table

When something feels wrong, check the surface before you rewrite the prompt.

Where Work and Codex overlap

OpenAI says Work and Codex have overlapping capabilities. If you prefer Codex, you can use it for research, documents, presentations, and other knowledge work. The difference is the interface and how the agent presents its work: Codex shows technical detail such as Git and shell commands, diffs, and review views; Work prefers nontechnical language and finished outputs.

That overlap is useful, but it is not a reason to use one for everything:

• Use Work when the person receiving the result does not want to see the machinery.
• Use Codex when the machinery is the point: the diff, the test, the command, the repository.
• Use Chat when the back-and-forth is the point and there is no file to hand over.

What the mode does not change

A mode changes the tools, the context, the history, and the output surface. It does not guarantee a better answer.

If the answer is vague in Chat, switching to Work will not fix the brief. If the code review misses the bug in Codex, switching to Chat will not make it appear. The mode is a room; the brief is still the work.

The one exception is context: Work can hold a larger task and use files and connected apps; Codex can see the repository and run commands. If the task needs that context, the mode is not cosmetic. It is the difference between having the evidence and guessing.

A 30-second chooser

Before you type, answer these in order:

1. Is the output a conversation or a file? Conversation → Chat. File → Work.
2. Does the task need a repository, terminal, tests, or a diff? Yes → Codex.
3. Does it need to run for a long time, pause for approval, or use connected files? Yes → Work.
4. Do I need to speak rather than type? Yes → Voice on a surface that supports it.
5. Am I in the right app? Codex is desktop; Work is web/mobile/desktop on eligible plans; Chat is everywhere.

If you cannot answer the first question, start in Chat. It is cheaper to discover the shape of the task there and move to Work when you know what the deliverable is.

Cost, without prices

Work follows the same usage structure as Codex and shares usage with it. Chat remains available on every configuration. We do not publish prices or limits because they change; the official pricing page is the current source.

A practical rule: if you are only asking a question, do not start a Work task just because it sounds more serious. If you need a reviewable file, do not keep re-prompting Chat and then copy the pieces into a document by hand. Pick the surface that matches the artifact.

If you see both ChatGPT and ChatGPT Classic

The new desktop app includes Chat and Work under ChatGPT, plus Codex. The previous desktop app may remain installed as ChatGPT Classic. It is not a third mode inside the new app; it is a separate app that continues to receive model updates, bug fixes, security patches, and support for its existing Enterprise capabilities. New agent features may be available only in the new app.

If your history looks split, check which app you are in before you assume something was deleted.

The short version

Chat for thinking, Work for deliverables, Codex for code, Voice when speaking is easier. Pick the room by the artifact, check the surface before blaming the prompt, and remember that Work and Codex share usage while Codex keeps its own history.

That is a smaller promise than “one mode to rule them all”. It is also the one that makes the other guides correct: most “the model is dumb today” reports are a mode mismatch, not a model failure.