ChatGPT Chat vs Work: What's the Difference and When to Use Each

The difference in one minute

ChatGPT now presents two main ways of working. Chat is the familiar conversation: you ask, it answers, you react. Work, which OpenAI launched on 9 July 2026, is described as an agent. You give it an outcome and it breaks the job into steps, pulls in information from files and connected apps, uses tools and produces something you can open afterwards, such as a document, spreadsheet, presentation or simple web app. OpenAI says it can keep going on a project for hours.

The practical split is who holds the steering wheel. In Chat you steer on every reply. In Work you steer at the start, at any checkpoints or approvals, and when you review the result. That makes Work faster for a well-defined job and a worse fit for a vague one, because a vague goal simply produces a confident-looking guess at length.

A third entry point, Codex, is for software development. It is not covered here.

Which one to pick for the job in front of you

The simplest test is to picture the finished result. If it is an answer or an idea, you want Chat. If it is a file or a piece of work someone else will open, Work is worth considering.

• Chat suits questions, explanations, brainstorming, rewording an email, summarising something you paste in and thinking a problem through when you do not yet know what you want.
• Work suits tasks with a clear outcome: a briefing pack from several documents, a spreadsheet built and checked from source data, a recurring weekly update or a first-draft deck for a known audience.
• Chat is usually the better choice for anything personal, sensitive or half-formed. Nothing about it needs access to your files or accounts to be useful.
• Work is the better choice when the job would otherwise be a dull hour of opening files, copying figures and formatting. The more precisely you can say what 'done' looks like, the better it tends to go.
• If you cannot describe the finished result yet, start in Chat to shape the brief, then hand the brief to Work. Going the other way, using Work for a quick factual question, mostly burns time and allowance.

What Work can touch that Chat usually does not

This is where the choice stops being about convenience. Work can use connected apps and, on the desktop app, work with files and programs on your computer with your permission. OpenAI's help material describes separate local and cloud ways of running tasks. Local tasks run on your machine, although messages and task context can still be stored in the cloud. Cloud tasks sync across web, mobile and desktop. Work can also run once, on a schedule, or in response to events in connected apps.

Because Work reads material and can then act on it, the main extra risk is untrusted content. A web page, shared document or incoming email can contain text written to steer an AI agent, which is called prompt injection. Security researchers describe the dangerous combination as private data, untrusted content and a route to send information out. Work can have all three at once. That is our reading of how the product is described rather than a confirmed incident, and OpenAI does not claim to eliminate the risk, so the sensible response is to limit what any one task can reach.

In practice that means a few habits:

• Start with read-only access to connected apps, and add permission to send, edit or delete only when you have a reason.
• Be wary of mixing client or financial data with open-ended web browsing in the same task.
• Treat anything Work sends, publishes, overwrites or deletes as needing your approval first. Reports of OpenAI's guidance say confirmation can be required for such actions and connector permissions can be adjusted, but check the settings on your own account rather than assuming.
• Check figures, dates, formulas and sources in whatever it produces. A tidy spreadsheet is not the same as a correct one.
• On Business, Enterprise and Edu plans, workspace administrators can control whether Work is available, including local and cloud access separately. If your organisation has a policy for AI tools, check it before connecting company accounts.

Plans, limits and what you will actually see

Availability depends on your plan, platform and, for organisations, your admin settings. At launch OpenAI said Work was on all plans on desktop, with web and mobile rolling out to Plus, Pro, Business, Enterprise and Edu. OpenAI's help material says desktop availability can still depend on plan and workspace, so what you see may differ from a colleague on another plan. On desktop, a toggle switches between the two experiences. Work is also not available in projects that use project-only memory.

On limits, the reliable part is the direction: long, multi-step Work tasks use more of your allowance than ordinary chat, and reporting suggests Work shares an allowance with Codex. The exact numbers are changing and third-party figures disagree, so we have not repeated them. Your plan's own help and pricing pages are the place to check. The same goes for which underlying model each mode uses; sources conflict on this, and it changes often.

The split may not last

In October 2026 it was reported that OpenAI's head of ChatGPT and Codex said on a podcast that Chat and Work will be merged into one experience. That comes from secondhand coverage of the podcast clip, not an OpenAI help page, and there is no published date. Nothing has been published on how a merged product would count usage.

The habit is likely to survive even if the buttons do not: decide whether you want an answer or a finished piece of work, and decide how much access the task needs. We will update this note if OpenAI documents the change.