What is agent mode in ChatGPT?
ChatGPT is now three experiences: Chat for conversation, Work for agentic tasks, and Codex for software development. Work is OpenAI's own description of "an agent designed for longer, multi-step work and finished deliverables" — it researches, analyses, works across connected apps and files, and produces documents, spreadsheets, presentations, reports and Sites. It runs on paid plans except Free and Go, and is metered like Codex rather than like chat.
- Current name
- Work alongside Chat and Codex
- What it produces
- Deliverables docs, sheets, decks, reports
- Plans
- Paid, not Free or Go Pro and Enterprise first
- Metering
- Like Codex shared agentic usage pool
- Local files
- Desktop only with your permission
- Scheduling
- Yes once, repeating, or triggered
- Codex on web
- Not selectable desktop app only
- Verified
- 2026-08-21 OpenAI help centre
The feature you searched for now has a different name and a different bill.
OpenAI's current documentation splits the product into three: Chat ("fast, conversational assistance and everyday questions"), Work ("an agent designed for longer, multi-step work and finished deliverables"), and Codex ("dedicated to software development and technical work"). If you came here looking for agent mode, Work is the experience that does what you meant.
A caveat on terminology, because we would rather flag it than paper over it. Nowhere in the documentation we read does OpenAI state that "agent mode" was renamed to Work — and the term is still live elsewhere in its own help centre, where Lockdown Mode is described as restricting "live web browsing, deep research, agent mode, file downloads". So the honest position is: Work is the agentic experience the product now ships and documents, and "agent mode" survives as a description of the capability rather than as the name of a distinct mode you toggle.
The pricing detail is the one that catches people out. OpenAI states plainly that "Work follows the same usage structure as Codex" and points at the Codex pricing page for included usage and additional credits — adding that its examples are coding-based, so Work usage will vary by task. That means agentic work and coding work draw on the same shared pool, and a heavy day of one leaves less of the other. It is not metered like ordinary chat, which remains unlimited.
Availability, precisely: Work is on ChatGPT web and mobile for eligible paid plans and in the desktop app where included, and it rolled out to paid plans except Free and Go, with Pro, Pro Lite, Enterprise and Edu first. Enterprise and Edu workspaces got a two-week preview with Work off by default. Codex, by contrast, is not selectable on web or mobile at all — desktop only, with remote access from the mobile app's Remote tab.
What it actually does for you.
The distinction between Chat and Work is the shape of the output. Chat gives you a reply; Work gives you a file. OpenAI's own framing for choosing between them: use Chat to ask, search or brainstorm; use Work to "research a topic, analyze information, or create a document, spreadsheet, presentation, report, or Site".
Practically, it also changes how you brief it. Work expects an outcome, the files and constraints it should use, and the criteria you will judge it by — and you can follow its progress, answer its questions, change direction and approve important actions while it runs. That is a different interaction from typing a prompt and reading a paragraph.
Work on web and mobile runs in the cloud and cannot reach files on your computer. In the desktop app, when your plan and workspace allow it, Work can use local files and desktop applications with your permission — and OpenAI notes that local files and outputs stay on that computer unless you move or share them.
Cloud Work chats sync across web, mobile and desktop; local chats stay on the machine. Workspace admins control the two independently: Work Cloud and Work Local are separate permissions, and Codex Local is a third. So "I have Work but it cannot see my files" is usually a permissions state rather than a bug.
Work can run once, repeat on a schedule or a trigger, or monitor for changes. That turns it from something you ask into something that keeps running — a weekly report, a watch on a source, a recurring cleanup.
Two things follow from that, and they are worth thinking about before setting one up. Anything recurring consumes the same shared agentic allowance every time it fires, without you present to notice. And anything monitoring an external source is exposed to whatever that source says — which is precisely the prompt-injection surface that Lockdown Mode exists to shut down, since it restricts browsing, deep research, agent mode and downloads together.
Part of this cluster asks the categorical question, and the accurate answer is: parts of it are. Chat is not an agent — it answers. Work and Codex are agentic in the meaningful sense: they take a goal, plan multiple steps, use tools, act on files and systems, and come back with a result rather than a suggestion.
Which is why the same product can be described as "just a chatbot" and "an agent that does your work" by two people who are both right, depending which toggle they are looking at. If the agentic half is what you want, note that it is not available on the free tier at all — the answer to "how do I use agent mode" begins with a paid plan above Go.
Where agentic work is actually priced.
Work is metered like Codex, so the honest comparison is against other agentic tools rather than against chat subscriptions. Prices from our verified plan sheet, checked 2026-07-16.
Frequently asked.
Quick follow-ups people search after this question.