ChatGPT agent mode is OpenAI's built in agentic feature that gives ChatGPT its own sandboxed browser and computer, so it can complete a multi step task instead of just explaining how to do it. Turn it on and ChatGPT browses real websites, fills in forms, pulls data from connected apps like Gmail or Google Drive, and hands back a finished document, spreadsheet, or answer, checking in only when a step needs your approval. Regular ChatGPT talks about a task. Agent mode goes and does it.
ChatGPT Agent Mode: Key Facts
- OpenAI launched ChatGPT agent on July 17, 2025, merging Operator (web actions) and deep research (source synthesis) into one system, rolled out first to Pro, Plus, and Team subscribers (OpenAI, "Introducing ChatGPT agent," and TechCrunch, July 17, 2025).
- The model behind agent mode scored 41.6 percent on Humanity's Last Exam, roughly double the 20.3 percent that OpenAI's o3 scores with no tools, rising to 44.4 percent across parallel attempts (OpenAI, "Introducing ChatGPT agent," July 2025).
- Agent mode is capped at roughly 40 tasks a month on Plus and about 400 on Pro, and on July 9, 2026, OpenAI extended the approach into a workspace mode, ChatGPT Work, built on its GPT 5.6 models (OpenAI usage limits; Forbes and Bloomberg, July 9, 2026).
ChatGPT Agent Mode vs Regular ChatGPT vs a Custom Built AI Agent
| Dimension | ChatGPT Agent Mode | Regular ChatGPT | Custom Built AI Agent |
|---|---|---|---|
| Autonomy | Runs a task on its own, pausing only for approval | None, answers one message at a time | Full autonomy within your design, including scheduled runs |
| Tool access | Own browser, terminal, file tools, plus connected apps like Gmail and Drive | Text and image generation, web search for answers only | Any APIs, databases, and systems you choose to connect |
| Best for | Ad hoc research, form filling, one time multi step jobs | Writing, brainstorming, quick answers, editing | Repeatable workflows that must scale or meet compliance needs |
| Limitations | Monthly task caps, no extensions or local file access | Cannot take action outside the chat window | Costs engineering time and budget to maintain |
| Setup required | None, built into ChatGPT | None, default mode | A development project, typically weeks |
What Is ChatGPT Agent Mode, Exactly?
ChatGPT agent mode is a mode you switch on inside ChatGPT, from the tools menu or by typing /agent, that gives the model a sandboxed virtual computer of its own: a browser, a terminal, and file tools it operates step by step to finish a task. OpenAI built it by merging two products it had shipped separately: Operator, which clicked and navigated live web pages, and deep research, which read dozens of sources and wrote synthesized reports. Agent mode combines both with ChatGPT's reasoning, so the same system that holds a conversation with you can also open a browser tab and finish what you asked for.
OpenAI treats agent mode as a step toward something bigger than one feature. Thibault Sottiaux, who now leads product across ChatGPT and Codex at OpenAI, told the Financial Times, "What we're building towards is where you have your own personal agent that is capable of helping you across everything in your life, be it personally or at work."
How Does ChatGPT Agent Mode Work?
ChatGPT agent mode works by opening a sandboxed virtual computer for each task: a browser, a terminal, and file tools that exist only for that session. You describe a task, agent mode plans the steps it needs, then executes them one at a time, checking the browser after each action before deciding what comes next.
It can type into search boxes, click links and buttons, scroll, fill out forms, and read the resulting page much as a person would, plus call connected apps, such as Gmail, Drive, or GitHub, to act on your own data. It reports what it is doing in plain language, and you can pause it, take over, or stop it at any point. Before anything consequential, such as a purchase or a message sent on your behalf, it stops and asks you to confirm. When finished, it hands back a filled in form, a spreadsheet, a slide deck, or a written answer with its sources.
What Tasks Can ChatGPT Agent Mode Handle?
ChatGPT agent mode is built for jobs that need several steps across real websites or apps, not single factual questions. Typical tasks include comparison shopping across retailer sites, researching a topic across dozens of sources and assembling a report or slide deck, filling out and submitting web forms, checking flight or hotel availability across several dates, pulling data from a connected Gmail, Drive, or GitHub account into a spreadsheet, and monitoring whether a product is back in stock.
The pattern is the same each time: agent mode replaces the browsing, clicking, and copying a person would otherwise do by hand, and delivers the finished output rather than instructions for producing it yourself.
How Is ChatGPT Agent Mode Different From Regular ChatGPT?
The difference is action versus conversation. Regular ChatGPT answers the message in front of it, whether that means writing text, generating an image, or searching for facts, then stops and waits for your next step. Agent mode keeps going: given a goal, it opens its own browser, works through the steps, and only returns when it needs approval or the job is done.
This is the same split that separates generative AI from agentic AI more broadly. Regular ChatGPT is generative: you prompt it, it produces one output, and the interaction ends there. Agent mode is agentic: given an objective, it plans and executes a sequence of actions, exactly the pattern our generative AI vs agentic AI guide covers in more depth. Regular ChatGPT cannot act outside the chat window, while agent mode logs into connected accounts and acts there, which is why it ships with far more safety guardrails.
What Are the Limitations of ChatGPT Agent Mode?
ChatGPT agent mode is capped, cautious, and sometimes slow by design. OpenAI limits agent tasks to roughly 40 a month on Plus and about 400 on Pro, resetting monthly, and a single task typically takes a few minutes to around thirty minutes, longer for complex jobs.
Inside the sandbox, it cannot install browser extensions, download files to your device, or reach other apps and files on your computer, and it pauses for confirmation before anything consequential, including purchases and messages sent on your behalf. OpenAI has also treated the feature with unusual caution, classifying it under the High level of its Preparedness Framework for biological and chemical risk, and disabling memory during agent sessions to limit prompt injection risk. It suits bounded jobs well, but can still misread a page or need a correction, so its output is worth checking, especially for anything customer facing or financial.
Is ChatGPT Agent Mode Free, or Only for Paid Plans?
ChatGPT agent mode is not available on the free plan. It shipped first to Pro subscribers, then to Plus and Team within days, with Enterprise and Education following soon after, and that paid only structure has held since. Plus gets roughly 40 agent tasks a month, Pro gets about 400, both resetting at the start of each calendar month regardless of when you subscribed.
If you are on the free tier and want agent style behavior, the options are upgrading to a paid plan, or looking at what a custom built agent can do instead, since that route is not gated by ChatGPT's plan tiers.
How Does ChatGPT Agent Mode Compare to Building a Custom AI Agent?
ChatGPT agent mode is the fastest way to get agentic behavior with no setup at all, while a custom built AI agent is slower to stand up but built around your exact workflow. Agent mode works inside OpenAI's sandbox, bound by its monthly task caps and the tools OpenAI supports. A custom agent is scoped and governed by the business that builds it: it connects directly to internal databases and proprietary APIs, runs on a schedule with no monthly ceiling, and applies the approval rules the business needs.
For a business process that needs more control and reliability than a consumer tool offers, custom AI agent development builds an agent scoped exactly to your workflow, connected to your own systems, and governed by rules your business sets. The tradeoff is straightforward: agent mode costs nothing extra and starts in seconds, while a custom agent costs engineering time and typically weeks to build, in exchange for removing the caps and generic guardrails of a consumer product.
Frequently asked questions
Is ChatGPT agent mode free?
No. ChatGPT agent mode is available only on paid plans: Plus, Pro, Team, Enterprise, and Education. Plus gets roughly 40 agent tasks a month and Pro gets about 400, both resetting monthly, with no free tier allocation.
Can ChatGPT agent mode browse the web and complete tasks on its own?
Yes. Agent mode opens its own sandboxed browser, navigates real websites, clicks, types, and fills out forms, then carries that through to a finished output such as a report or completed form, pausing only when it needs your confirmation.
Is ChatGPT agent mode safe to use for business tasks?
It is reasonably safe for bounded tasks thanks to guardrails like confirmation prompts and restrictions on extensions or local file access, though OpenAI recommends watching its work on sensitive sites. For payments or compliance heavy workflows, a custom built agent is generally the safer choice.
How long does a ChatGPT agent mode task take to finish?
Most tasks take a few minutes to about thirty minutes, though complex jobs touching several websites can take longer. You can watch its progress in real time and stop or take over at any point.
What is the difference between ChatGPT agent mode and Operator?
Operator was OpenAI's earlier standalone tool for taking browser actions. On July 17, 2025, OpenAI folded Operator together with deep research into one system, ChatGPT agent, so Operator no longer exists as a separate product.
How do I turn on ChatGPT agent mode?
On a Plus, Pro, Team, Enterprise, or Education plan, open the tools menu in the composer and select agent mode, or type /agent followed by your task. ChatGPT will begin planning and executing the steps, checking in when it needs input.
Updated July 2026. ChatGPT agent mode's task caps, connected apps, and available plans change as OpenAI ships updates, most recently the July 9, 2026 introduction of the GPT 5.6 powered ChatGPT Work mode, so confirm current limits inside ChatGPT's settings before building a workflow around them.
Need an agent built around your own workflow, not a consumer tool's limits? Book a free scoping call with Codioo's AI agent team.