Product images: OpenAI, Google Cloud and Anthropic official pages
I have worked in IT for about twenty years, and no word has been stretched as far in the last two years as "agent". When someone asks whether ChatGPT is an AI agent, the honest answer depends on which part of ChatGPT they mean, and on whose definition of an agent they use. This guide follows the searches people run next on Google (agent or LLM, agent or chatbot, is it free, Claude, examples, agent mode, 2022) and answers each from OpenAI, Anthropic and Google Cloud's own pages, all checked on October 28, 2026.
Quick answer- Plain chat: a conversational assistant. OpenAI's Work and Codex help page describes Chat as fast, conversational assistance.
- ChatGPT Work: OpenAI calls it "an agent designed for longer, multi-step work and finished deliverables", so yes, that part is an agent.
- LLM or agent: the LLM is the model inside. The agent is the model plus tools, instructions and permission to act.
- Free: no. OpenAI's ChatGPT agent page now says the original agent mode is no longer available and points to Work, which needs an eligible paid plan.
- Claude: same rule. Anthropic's agents post says the model must direct its own process and tools.
Jump to: Short answer | What is an AI agent | Agent or LLM | Agent or chatbot | Agent mode | Is it free | Is Claude an agent | Examples | 2022 | Safety | Five-question test | FAQ
Product names, plans and limits change quickly. Treat every plan detail as a snapshot and check the linked page before you pay.
Is ChatGPT an AI Agent? The Short Answer
Image: OpenAI (official Help Center page)
The question has a moving answer, and OpenAI's own help center shows why. The page for the original ChatGPT agent now opens with "ChatGPT agent is no longer available" and sends readers to ChatGPT Work for "longer, multi-step tasks and finished deliverables", and to a cloud browser for supported browser workflows. The product did not disappear. The agent idea moved into a different part of the app.
The Work and Codex page explains the new split. ChatGPT includes Chat and Work, and Codex sits in the desktop app. Chat is for "fast, conversational assistance and everyday questions". Work is "an agent designed for longer, multi-step work and finished deliverables". Codex "remains dedicated to software development and technical work". So one subscription holds an assistant and an agent side by side.
Image: OpenAI (official Help Center page)
| Part of ChatGPT | How OpenAI describes it | Agent or assistant? |
|---|
| Chat | Ask questions, search, brainstorm, or get quick conversational help | Assistant. You steer every turn. |
| Work | Research a topic, analyze information, or create a document, spreadsheet, presentation, report, or Site | Agent, in OpenAI's own words |
| Codex | Write or debug code, run tests and commands, review changes, or work with a repository | Agent-style, for software work |
| Cloud browser inside Work | Its own browser on a separate computer in the cloud that can read pages, click buttons and fill forms | A tool the agent uses |
My take (20 years in IT): When a vendor says "agent", ask what it can do without you pressing another key. If the answer is "write text for me to paste", it is an assistant. If the answer is "open the site, fill the form and stop before paying", it is an agent, and it needs the same care you would give a new employee with your login.
What Is an AI Agent? Three Official Definitions
There is no single agreed definition, and the three big vendors phrase it differently. Reading all three side by side is the fastest way to stop arguing about the word.
Image: OpenAI (official guide page)
OpenAI. Its practical guide to building agents (January 15, 2026) says "Agents are systems that independently accomplish tasks on your behalf." An agent uses an LLM to manage workflow execution and make decisions, and it has tools to gather context and take actions within guardrails. The guide is blunt about the other side: applications that integrate LLMs but do not use them to control workflow execution, "think simple chatbots, single-turn LLMs, or sentiment classifiers", are not agents.
Image: Anthropic (official page)
Anthropic. Its building effective agents post groups everything under "agentic systems" and then draws one line. Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents are systems where LLMs dynamically direct their own processes and tool usage. A page note says much of the tooling described has changed since December 2024, but the distinction is still the clearest one I have seen.
Image: Google Cloud (official page)
Google Cloud. Its what are AI agents page defines agents as software systems that use AI to pursue goals and complete tasks on behalf of users, showing reasoning, planning and memory and a level of autonomy. It lists reasoning, acting, observing, planning, collaborating and self-refining as key features.
All three agree on the heart of it. An agent decides the next step itself and uses tools to act, and a chat window that only returns text does neither.
Is ChatGPT an AI Agent or LLM?
This question mixes two layers. An LLM is a model, a thing that predicts and generates language. ChatGPT is a product built around models. OpenAI's guide says an agent has three core components, the model, the tools and the instructions, and Google Cloud calls the LLM the "brain" of an agent. So the LLM is a part, and "agent" describes the whole system.
Image: OpenAI (official guide page)
The Work page makes the product-versus-model difference concrete. It lists models such as GPT-6.1 Sol and GPT-6 Luna as models for ChatGPT Work and Codex and says they are "not available in regular ChatGPT conversations". The same account therefore runs different models for different jobs, which would not make sense if ChatGPT were just one LLM.
Image: Google Cloud (official page)
Image: Anthropic (official page)
Anthropic's version is the augmented LLM, a model enhanced with retrieval, tools and memory, which it calls the basic building block of agentic systems. That is a useful picture for ChatGPT. The bare model is the LLM, the augmented model is what you get in Chat when it searches or uses connected apps, and the agent is the loop that keeps going until the job is done.
Is ChatGPT an AI Agent or Chatbot?
Google Cloud's page has a table that settles most of this. It separates agents, assistants and bots by purpose, interaction and decision-making. An assistant is reactive and responds to user requests, and it "can recommend actions but decision-making is done by the user". An agent is proactive and goal-oriented, and it can make decisions independently. A bot follows pre-defined rules.
Image: Google Cloud (official page)
By that table, ChatGPT Chat is an assistant. It is more capable than a rule-based bot, but you still choose each step. Work moves toward the agent column because you describe an outcome, it plans and carries out the steps, and it pauses for clarification or approval. One caution from the Work page itself: Chat can "search", so tool use alone is not the dividing line. The difference is who drives the loop.
Image: OpenAI (official Help Center page)
ChatGPT Agent Mode and What Replaced It
Agent mode was the first place ChatGPT clearly acted as an agent. The help page says you started it from the tools menu or by typing /agent, described the task and let it run, and it would "pause for clarification or confirmation when needed". It used screenshots of its virtual browser to see pages and click buttons. The page also says Operator functionality was integrated into agent mode and that the Operator website is no longer accessible.
Image: OpenAI (official Help Center page)
Today the same page says the agent is gone and sends you to Work. The browsing part continues as cloud browser, which gives Work "its own browser on a separate computer in the cloud" that can read pages, click, enter information in forms and work on supported public and signed-in websites. You do not pick it separately, because ChatGPT chooses when to use it, or may use a connected app or plugin instead.
Image: OpenAI (official Help Center page)
Image: OpenAI (official Help Center page)
Three controls matter. By default ChatGPT asks before it opens a new website, and Settings has three modes: Always ask, Auto approve, and Always allow, which OpenAI says is not recommended. It is designed to ask for confirmation before actions that could be hard to reverse or create a financial, legal or account commitment, such as confirming a booking or making a payment. And if it gets stuck, it asks you to take over the browser.
Image: OpenAI (official Help Center page)
Work can also run on its own schedule. Scheduled tasks can run once, repeat, monitor for changes, or respond to supported events. Event-triggered tasks run in Work and respond to Gmail, Slack or GitHub activity, and they are not available on Free or Go.
Image: OpenAI (official Help Center page)
Is ChatGPT Agent Free?
Not in the way the search implies. The help page for the original agent mode says it was available only on paid plans, and lists monthly message limits of 40 for Plus, 400 for Pro and 40 for Business and Enterprise, with 30 credits per message on flexible pricing. Because that page now says the agent is no longer available, treat those numbers as history. The current agent experience is Work, and that follows the plan rules below.
| Feature | Free | Paid plans | Source |
|---|
| Work (agent for longer tasks) | Not listed | Available on eligible paid plans | Work and Codex page |
| Cloud browser | Excluded (Free and Go) | Available in Work in supported regions | Cloud browser page |
| Scheduled tasks | Yes, 3 active, one-time or at most daily | 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise | Scheduled tasks page |
| Event-triggered tasks | No | Eligible Plus, Pro, Business, Enterprise, Edu | Scheduled tasks page |
Image: OpenAI (official Help Center page)
Two cautions. Availability depends on your region and workspace settings, and OpenAI says Work usage "will vary by task" because it draws from the same usage structure as Codex. If you plan to rely on it, read the Work page's usage section on the day you buy, not the day you read a review.
Is Claude an AI Agent?
The same test applies. Anthropic's post says agents are "typically just LLMs using tools based on environmental feedback in a loop", and that they plan and operate independently, pausing for human feedback at checkpoints or blockers. Whether a Claude product is an agent depends on whether the model is running that loop, not on the brand name.
Image: Anthropic (official Claude Code Docs page)
The Claude Code overview is the clearest example I found. It calls Claude Code an "AI-powered coding assistant" and then describes behavior that matches the agent definitions. You describe what you want, and Claude Code "plans the approach, writes the code across multiple files, and verifies it works". It can spawn multiple agents, where a lead agent coordinates the work and merges results, and an Agent SDK lets you build custom agents. I did not find an Anthropic page that labels an ordinary chat reply as an agent.
Image: Anthropic (official Claude Code Docs page)
ChatGPT Agent Examples You Can Check Yourself
OpenAI's cloud browser page lists the kinds of tasks it has in mind, and they are a good reality check because each one has several steps and a site you do not control.
- Check restaurant availability or request quotes from businesses.
- Find flights that match your plans, or compare product availability and local stock.
- Track a package using details from a connected email account and a tracking website.
- Find a DMV appointment and prepare a booking for your approval.
The page adds that what ChatGPT can complete depends on the website, your access and the steps involved, and that some tasks need your review or a final step you complete yourself. The simplest scheduled example on the scheduled tasks page is asking it to let you know when your package is delivered.
Outside ChatGPT, Anthropic's two production examples are customer support and coding. Support works as an agent use case because conversations need both talking and actions such as updating tickets, and coding works because results can be checked by tests. For business setups, see my guide to AI agents for business operations.
Image: Anthropic (official page)
Image: Anthropic (official page)
Image: Google Cloud (official page)
Many things sold as "agents" are really the workflows in Anthropic's diagrams, a fixed chain of model calls or a router that sends each request down one path. They are useful, cheaper and more predictable, and they are not agents in Anthropic's sense. When you evaluate a product, ask whether the path is fixed or whether the model chooses it.
My take (20 years in IT): If a task is the same every time, build it as a workflow and keep the model on a short leash. Use an agent only when the steps genuinely vary. In my experience the failures I see in agent projects come from using autonomy where a checklist would have worked.
Is ChatGPT an AI Agent in 2022?
People search this phrase when they remember the first version of ChatGPT as a text box that answered questions. I did not find an OpenAI, Anthropic or Google Cloud page describing a plain chat exchange as an agent, and the definitions above would call it a chatbot, since it returned text and the person decided the next step.
What changed is the set of agent-style features added around the model. Today's help pages describe agent mode with a /agent command, Work, a cloud browser, scheduled and event-triggered tasks, and apps connected to Gmail, Slack and GitHub. If you saw a 2022 article calling ChatGPT an agent, check what it meant by the word, because those features are described in OpenAI's current documentation, and I cannot date each one from these pages.
What Changes When ChatGPT Acts for You
An assistant can only be wrong. An agent can be wrong and then do something. OpenAI's agent page says that when you sign it into websites or enable apps, it can access sensitive data such as emails, files and account settings and perform actions like sharing files or changing settings, which creates privacy risks including prompt injection.
Image: OpenAI (official Help Center page)
The page gives a vivid example. You ask the agent to find a restaurant by checking your calendar and recent emails. While researching, it meets a malicious comment that tells it to retrieve a password reset code from Gmail and send it to a malicious website. OpenAI lists safeguards, including user confirmations for high-impact actions, refusal patterns, prompt injection monitoring and a watch mode that needs your supervision on certain sites. It adds that "these measures don't eliminate all risks".
Image: OpenAI (official Help Center page)
On data, the page says Business, Enterprise and Edu data is not used for training by default, while for Plus and Pro, data including screenshots is used under OpenAI's privacy policy, and you can turn off "Improve the model for everyone" in Data Controls. Deleted chats and their screenshots are removed from OpenAI's systems within 90 days.
Image: OpenAI (official Help Center page)
Image: OpenAI (official guide page)
Anthropic gives the sober version of the same advice: find the simplest solution possible and add complexity only when needed, because agentic systems often trade latency and cost for better task performance, and autonomy means higher costs and the potential for compounding errors. It recommends testing in sandboxed environments with guardrails.
A Five-Question Test You Can Use on Any Product
This checklist is my own, built from the three definitions. It is a way to organize questions, not a certification.
- Does it choose its own next step? If you pick every step, it is an assistant.
- Does it use tools? Browsing, files, apps and code are tools. Plain text alone is not.
- Does it act outside the chat box? Sending, booking, editing and paying are actions.
- Does it keep going after your first prompt? Long tasks and schedules point to an agent.
- Does it stop before risky steps? A good agent asks first.
Chat answers yes to question 2 sometimes and no to most of the rest. Work and the cloud browser answer yes to most, and OpenAI's pages say it asks for confirmation on consequential actions. Claude Code, by Anthropic's own description, answers yes to the first four. For a decision on what to buy, the answer to question 5 matters more than the answer to question 1.
My take (20 years in IT): Before you give any agent a login, write down the worst thing it could do with that account and decide whether you would let a stranger do it for one afternoon. If the answer is no, use a separate account with limited rights, or keep that task for yourself.
Frequently Asked Questions
What is the difference between ChatGPT and an AI agent?
A chat assistant answers when you prompt it and leaves the next step to you. An agent, in OpenAI's words, is a system that independently accomplishes tasks on your behalf, using a model, tools and instructions. In ChatGPT, OpenAI describes Chat as fast, conversational assistance and Work as an agent for longer, multi-step work and finished deliverables.
Is ChatGPT an AI agent or an LLM?
Neither word fits the whole product. The LLM is the model inside. ChatGPT is the product around it, and the Work and Codex experiences add tools and permissions. OpenAI's help page lists some models that are available in Work and Codex but not in regular ChatGPT conversations.
Is ChatGPT agent free?
No. OpenAI's help page says the original agent mode was only for paid plans and now says ChatGPT agent is no longer available, pointing to ChatGPT Work. Work is available on eligible paid plans, and the cloud browser is excluded from Free and Go. Free accounts can still create basic scheduled tasks within limits.
Is Claude an AI agent?
It depends on how it is used. Anthropic separates workflows, where code controls the steps, from agents, where the model directs its own process and tools. Anthropic's Claude Code docs describe a tool that plans, edits files and runs commands, and an Agent SDK for building your own agents. A plain chat reply is not that.
Can I use ChatGPT to create an AI agent?
You can use ChatGPT to help design one, but building one means giving a model tools and instructions. OpenAI's practical guide names three parts, the model, the tools and the instructions, and recommends starting with a single agent before adding more.
What is ChatGPT agent and why is it on my website?
OpenAI says ChatGPT agent used a virtual browser and that it may not be able to visit certain websites. Its help page links an allowlisting guide for site owners who want it to reach their pages. Its cloud browser page adds that each website decides whether to allow automated browser traffic, so check your own bot and firewall rules.
Try These Tools
Official vendor pages. These are not affiliate links.
Sources
- OpenAI Help Center, ChatGPT agent (checked October 28, 2026)
- OpenAI Help Center, ChatGPT Work and Codex (checked October 28, 2026)
- OpenAI Help Center, Using cloud browser in ChatGPT (checked October 28, 2026)
- OpenAI Help Center, Scheduled tasks in ChatGPT (checked October 28, 2026)
- OpenAI, A practical guide to building agents (January 15, 2026) (checked October 28, 2026)
- Anthropic, Building effective agents (checked October 28, 2026)
- Google Cloud, What are AI agents (checked October 28, 2026)
- Claude Code Docs, Overview (checked October 28, 2026)
Product images come from the vendors' official pages, are cropped and framed here, and belong to their owners. I did not test these products myself, and claims about what they do are the vendors' own.
Last updated: October 28, 2026.
Comments
Post a Comment