Atlas and Codex, in short
ChatGPT Atlas is OpenAI's own web browser, launched in October 2025 and built on Chromium, with ChatGPT in a sidebar so it can read the page you are on, summarise it, compare things across tabs and, in agent mode, carry out tasks like filling forms.
Codex is OpenAI's coding agent. You describe what you want in plain English and it writes, edits and runs code across a whole project rather than handing back a single snippet.
- Atlas is for everyone. Codex is for people who write software.
- Both mark the shift from AI that answers questions to AI that does the task.
- Give an agent access the way you would give a new temp your logins: sparingly.
- OpenAI keeps folding Atlas features into the main ChatGPT app, so menus move.
What is ChatGPT Atlas?
ChatGPT Atlas is OpenAI's own web browser, launched in October 2025 and built on the same foundations as Chrome. The idea is simple: ChatGPT sits alongside every page you visit, so it can read what you are looking at, summarise it, compare things across tabs, and in "agent mode" actually do things.
How Atlas differs from Chrome with a ChatGPT tab open
The difference is context. A normal browser makes you the courier: you copy a page, paste it into a chat, ask your question, then carry the answer back. Atlas removes the courier. The assistant already sees the page, the tabs and the flow of what you are doing, so "summarise this", "compare these three" and "pull the key dates out of this" stop being copy-and-paste jobs.
What agent mode actually does
Agent mode is where it stops advising and starts acting: navigating pages, filling in fields, gathering a shortlist, working through a multi-step task while you watch. It is genuinely impressive and occasionally cheerfully wrong, which is exactly the combination that requires supervision.
Where Atlas earns its keep
- Research with a lot of tabs. Comparing suppliers, venues, specs or prices without twelve copy-pastes.
- Reading before deciding. Summarising a long page, a policy or a set of terms in the flow of work.
- Repetitive form-filling. The dull, structured jobs nobody has ever enjoyed doing by hand.
- Drafting in context. Writing a reply or a summary while the source is still on screen.
Is ChatGPT Atlas safe to use at work?
It is genuinely useful and genuinely worth being careful with. An assistant that can act inside your browser is powerful precisely because it can act inside your browser. It inherits your logged-in sessions, which is the whole point and the whole risk in one sentence.
Think in permissions, not vibes
Give it access the way you would give a new temp your logins: sparingly, and only for what the job needs. Nobody would hand a first-day temp the payroll system and pop out for lunch. Same instinct here.
Five rules before you let an agent into your browser
- Keep it out of money and payroll. Banking, invoicing and HR systems are not the place to test a new tool.
- Review before you confirm. Read what it has filled in before anything is submitted or sent.
- Separate profiles. Use a browser profile without your client and finance logins for anything agentic.
- Assume the page can lie to it. Content on a web page can try to influence an agent. Treat instructions found on a page as untrusted.
- Write it down. One line in your AI policy beats a conversation after something goes wrong.
One practical note: OpenAI has been folding Atlas and its agent features into the main ChatGPT app, so if the menus have moved by the time you read this, that is why. The direction of travel is clear even when the buttons keep shifting.
What is Codex?
If you have heard the word Codex and wondered whether you are missing something, you probably are not. Codex is OpenAI's coding agent, aimed at developers. You describe what you want built and it writes, edits and runs code across a whole project rather than handing back a single snippet you paste somewhere and pray.
Codex versus asking ChatGPT for code
The ordinary chat box is a very clever colleague who has never seen your codebase. It gives you a snippet, and you do the integrating, the testing and the swearing. Codex works inside the project: it reads the existing files, changes several of them, runs the thing and iterates on what comes back. That is the difference between advice and work.
Who Codex is actually for
Developers and technical teams. For most people reading this, it is a tool for another day. But it is worth knowing it exists, because it shows where all of this is going: less "AI that answers questions", more "AI that does the task while you supervise". If your team writes software, Codex and tools like it are already changing how the work gets done, and someone in the building should understand them properly.
The same shift is happening on the Anthropic side with Claude Code, which we teach directly in Introduction to Claude Code, Advanced Claude Code and Building Agents in Claude Code. The concepts transfer: if your people understand one agentic tool, they understand the category.
Atlas, Codex and the shift from answers to actions
Put the two together and the pattern is obvious. For three years the question was "what can it tell me?" The question now is "what can it do, and how much of it do I let it do unsupervised?" That is a management question as much as a technical one, and it lands on teams that are still getting comfortable with writing a decent prompt.
What this means for your team, practically
- Someone needs to own it. One named person who understands what agents can reach inside your business.
- Checking becomes the skill. Reviewing an agent's work well is harder, and more valuable, than prompting.
- Permissions are policy. Decide in advance which systems an agent may touch, and write it down.
- Start with low-stakes tasks. Research, summarising and form-filling before anything that spends money.
None of this requires a technical team. It requires the same thing every other AI decision requires: a bit of plain-English understanding, spread across enough people that the decisions get made deliberately rather than by whoever downloaded something on a Friday afternoon.
ChatGPT Atlas and Codex, Explained in Plain English: frequently asked questions
What is ChatGPT Atlas?
ChatGPT Atlas is OpenAI’s web browser, launched in October 2025 and built on Chromium, with ChatGPT sitting in a sidebar so it can read the page you are on, summarise it, and, in agent mode, carry out tasks like filling forms or pulling together a shortlist. OpenAI has since been folding Atlas and its agent features into the main ChatGPT app, so check the current state before you rely on any single version.
Is ChatGPT Atlas free?
Atlas itself is a free download, but what you can do inside it depends on your ChatGPT plan. The heavier agent features have historically been tied to paid tiers and usage limits. Treat the browser as free and the capability as part of your subscription.
Is ChatGPT Atlas safe to use at work?
It is safe enough to use deliberately and risky to use casually. An assistant that can act inside your browser is powerful precisely because it can act inside your browser, so it inherits your logged-in sessions. Keep it out of banking, payroll and client systems, review anything it does before you confirm, and set a rule for what staff may let it touch.
What is agent mode in ChatGPT?
Agent mode is where ChatGPT stops describing what to do and starts doing it: navigating pages, filling in fields, gathering information across tabs and working through a multi-step task while you watch. You stay the supervisor. The useful mental model is a fast, literal-minded intern who never asks whether this seems like a good idea.
What is Codex?
Codex is OpenAI’s coding agent. You describe what you want in plain English and it writes, edits and runs code across a whole project rather than handing back a single snippet. It is aimed at developers and technical teams, not first-time users, and it is a separate tool from the ordinary ChatGPT chat box.
What is the difference between Codex and asking ChatGPT for code?
The chat box gives you a snippet to copy and paste, and it cannot see the rest of your project. Codex works inside the project: it reads the existing files, makes changes across several of them, runs the code and iterates on the result. That is the difference between advice and work.
Do I need to learn Codex if I do not write code?
No. It is worth knowing the word so you are not lost in a meeting, and that is all. If your organisation ships software, however, someone in the building should understand it properly, because it changes how the work gets estimated and reviewed.
Are AI browsers like Atlas replacing normal browsers?
Not yet, and probably not in the way the headlines suggest. What is clearly happening is that assistants are moving into the places where the work already happens: your browser, your editor, your documents. Whether that arrives as a separate browser or as a feature inside the ChatGPT app matters far less than getting your permissions and habits right.
Harry Lang runs The Oxford AI School, teaching teams and businesses the basics of AI in a fun, easy and practical way. A marketing leader turned AI trainer, he can be reached at Learn@TheOxfordAISchool.com.
Practical ChatGPT training for you or your team
The Oxford AI School delivers plain-English AI training in Oxfordshire and London, across the UK, and live online for teams in the USA and other international markets. Sessions cover prompting, checking output, data safety, images, agents and the tools your team already has open.
Where teams usually start
Basic AI Training
Learn prompting, checking, data safety and practical everyday use in a morning.
Explore this trainingAI Basics Deep Dive
Go further on prompting, projects, images and research with hands-on practice.
Explore this trainingWhich AI Tool to Start With
Compare ChatGPT with Claude, Gemini, Copilot and Perplexity for your business.
Explore this trainingAI Policy Workshop
Set clear rules for safe ChatGPT use before your team pastes something it shouldn't.
Explore this trainingThe ChatGPT series
Four plain-English guides to the questions people actually type into Google and ChatGPT about ChatGPT.
Get your team ready for AI that acts, not just answers
Agents change what your people need to know about permissions, checking and accountability. We cover exactly that, in plain English, on your own tools.
Book a call