You finished a first half hour on web or mobile. Ordinary chat works. Then the desktop app shows up with more doors, and the question becomes which door is Chat, which door is Work, and why Codex sits next to both like a different job entirely.
This is Part 4 of Learn ChatGPT from scratch. Part 3 got you through account, a real question, a safe file, and a rewrite. Here we stay on the desktop app as a product map, not a full Work tutorial and not a coding bootcamp. You will learn what Chat, Work, and Codex are for, what Free still limits, how usage burns differently, and how to choose a mode without turning every task into an agent experiment.
Model names on the Work side of OpenAI’s announcements have been moving quickly (including GPT-5.6-related Computer Use notes in Codex). Treat model labels as “what the UI shows today,” not as permanent brand names in your runbook. For the Anthropic sibling map (chat vs Code vs Cowork), see Learn Claude from scratch and the Claude product map series. Paths also sit on Learn.
What you will leave with
- Where to get the desktop app and how it relates to older ChatGPT Classic and Codex-only apps
- Plain definitions: Chat (conversation), Work (agentic multi-step office work), Codex (coding)
- Plan reality: Free can see more of the desktop, but Work and Codex are limited compared with paid use
- A mode chooser you can use before you start a long task
- Usage habits so multi-step modes do not empty your budget by accident
- What to practice this week without installing a full engineering workflow
Why desktop is a different lesson
On web and mobile, most people live in conversation. You type, you get a reply, you rewrite, you leave. The desktop app still does that. It also packages agent-style office work and coding work into the same install so you stop juggling three separate apps when OpenAI merges product lines.
OpenAI’s mid-2026 desktop direction is explicit: the ChatGPT desktop app for Mac and Windows is meant to hold Chat, Work, and Codex together. The older Codex app can update into this combined app. Older “ChatGPT Classic” desktop builds may still exist as a migration path while people move. UI labels shift; the three job types are the durable map.
If you only need email rewrites and short explanations, web is still enough. Desktop pays off when you want a dedicated window, local-ish work patterns (files, screen context, coding projects), and clearer mode separation than a single browser tab can show.
Install and land (without a rabbit hole)
- Open the official download hub at chatgpt.com/download (or the desktop feature page OpenAI links from help articles).
- Install for Mac or Windows from that path. Prefer the official page over random third-party mirrors.
- Sign in with the same account you used on web in Part 3.
- Look for the mode switcher (Chat / Work / Codex or equivalent labels). If your build only shows a subset, note what you have and keep reading; web Chat still covers conversation.
- If you already used a standalone Codex app, update it per OpenAI’s migration help. Existing Codex chats and projects are intended to carry over when the product merges cleanly.
You do not need every permission grant on day one. Folder access, browser control, and connector grants belong to later tutorials. Today is orientation: which mode for which job.
Three modes, three jobs
| Mode | What it is | Best for | Not the first stop for |
|---|---|---|---|
| Chat | Conversation: you lead turn by turn | Drafts, rewrites, Q&A, learning, one-file review | Long unsupervised multi-app office jobs |
| Work | Agentic office-style multi-step help | Delegated tasks that need several steps, tools, or deliverables you still approve | A single paragraph rewrite you could finish in Chat |
| Codex | Coding-focused agent surface | Repos, diffs, PRs, developer workflows | Stakeholder emails and pure prose research |
Chat: conversation
Chat is the mode you already practiced in Part 3. You own the loop: prompt, read, rewrite, export. Use it when the work is language-heavy, judgment-heavy, or small enough that an agent would just add ceremony.
Examples that stay in Chat:
- Rewrite a project update from bullet notes
- Explain a metric definition for non-analysts
- Turn a messy brief into three outline options
- Review one PDF you are allowed to share and list open questions
Work: agentic multi-step office work
Work is the “do more than one turn of typing” mode for non-coding knowledge work. Think multi-step: gather context, draft a longer artifact, use tools you allow, produce something closer to a finished package that you still review. It is the product shape people compare to Claude’s Cowork-style agent work, even when the UI labels and permissions differ.
Examples that belong in Work when your plan and permissions allow:
- Turn a folder of meeting notes into a structured brief with open questions
- Build a multi-section report from several sources you explicitly provide
- Run a longer delegated task where intermediate steps matter and you want progress, not one shot of text
Work is not a free pass to unsupervised action on company systems. Approvals, review, and data rules still sit with you. Later series parts in the ChatGPT Work tutorial track go deeper on first agentic tasks. Here you only need to recognize the mode and not waste it on a subject-line rewrite.
Codex: coding
Codex is the coding agent experience living inside the same desktop shell. It is for developers and technical professionals: repositories, diffs, pull request review side panels, multi-repo projects, and computer-use style flows where the product supports them. If your job is SQL drafts, Python scripts, or application code, Codex is the door. If your job is a stakeholder email, it is the wrong door with a higher chance of expensive side quests.
OpenAI’s public notes around the desktop merge mention coding workflow upgrades (inline edits in diffs, PR review in a side panel, multi-repository projects) and faster computer use powered by a current-generation model line that has been marketed alongside GPT-5.6 branding. Hedge that: your build may show a different model string next month. Read the name in the UI, do not hardcode last week’s blog post into your team wiki.
Plan reality: Free vs paid on desktop
Marketing lines often say Chat, Work, and Codex are available on every plan, including Free, on the desktop app. That can be true as a “you can open the doors” claim and still false as a “you get the same capacity” claim. Free is useful for orientation. Multi-step Work and Codex tend to burn usage faster and hit limits sooner. Plus, Pro, Business, and Enterprise are where heavier Work and coding patterns usually become practical day to day.
Mobile notes also matter for expectations: download pages have described ChatGPT Work on phones with paid plan language in places. Do not assume every desktop button has a full twin on Free mobile. When money or SLA is real, re-check OpenAI’s current plan matrix for your account type.
| Question | Practical answer |
|---|---|
| Can Free open the desktop app? | Yes, downloads are for global Mac and Windows use with a normal ChatGPT login. |
| Can Free touch Work or Codex? | Often yes as a limited preview of the modes; capacity is the constraint, not curiosity. |
| When should I pay? | When Chat limits or Work/Codex capacity block real weekly work, not because a banner shouted once. |
| Business / Enterprise? | Org admin, data controls, and workspace choice matter more than mode labels. Follow IT policy. |
Mode chooser: pick before you type
Ask three questions out loud before you switch modes:
- Is the main artifact language, or code? Language-first → Chat or Work. Code/repo-first → Codex.
- Is one careful conversation enough? If yes → Chat. If the job needs multi-step tool use and longer delegated progress → Work (or Codex for engineering).
- Am I willing to review intermediate steps? If you need a 30-second subject line, stay in Chat. Agents shine when review is part of the job, not when you wanted a faster autocomplete.
Quick scenarios
| You need… | Start in | Why |
|---|---|---|
| A shorter email from notes | Chat | One loop; agent overhead is waste |
| A multi-source brief with steps and files | Work | Multi-step office task |
| Explore a repo and propose a patch | Codex | Coding surface |
| Learn what a feature flag means | Chat | Q&A, not a project |
| Review a PR and comment on diffs | Codex | Built for that workflow |
| Turn three meeting PDFs into a plan | Work (if allowed) | Multi-step + deliverable |
If you pick wrong, stop early. Copy the useful text out, open the right mode, and start clean rather than dragging a coding agent through a press-release rewrite.
Usage burns faster on Work and Codex
Chat is usually the cheapest way to spend a limit: one prompt, one reply, a couple of follow-ups. Work and Codex spend more because multi-step agents take more tool calls, more context, and more “thinking” time. That is not a moral failing of the product. It is why Free feels generous for chat demos and tight for agent demos.
Habits that protect you:
- Scope the job in one sentence before you start Work or Codex. Vague agent goals burn tokens while the model explores.
- Prefer Chat for the first draft of language, then only move to Work if the remaining work is truly multi-step.
- Do not open Codex for prose. You will get tooling noise and usage burn for a job Chat already does.
- Watch plan usage indicators when the product exposes them. If you hit a wall mid-task, finish the export of what you already have instead of re-running the whole agent loop.
- On shared Business seats, heavy Work/Codex sessions can be a team cost, not a private hobby. Know your org’s norms.
Rule of thumb: if the task fits in a single screen of notes and a single deliverable paragraph set, stay in Chat. If the task needs a checklist of steps and tools, consider Work or Codex.
Worked example: same goal, three wrong and right doors
Goal: “Turn last week’s messy product notes into a one-page status for leadership, then list three open engineering questions.”
Wrong: Codex first
You open Codex because it looks powerful. The product expects a repo. You spend fifteen minutes explaining that there is no code. Usage burns. You get a partial markdown file that still needs the same rewrite Chat would have done faster.
Wrong: Work for a two-paragraph email
You launch Work for a short status. The agent asks for extra context, tools, and steps. You needed a draft and a human edit. Chat would have finished in two turns with less burn.
Right: Chat first, Work only if the job grows
Role: Draft for human review.
Audience: leadership + eng lead.
Goal: one-page status + 3 open engineering questions.
Tone: plain, no hype.
Constraints: 350-500 words. No invented metrics.
If a number is missing, say "unknown" instead of guessing.
Source notes:
[paste bullets]
Task:
1) Status draft with sections: shipped, in progress, blocked, next.
2) Three open eng questions (only from notes).
3) Assumptions list.
4) Unknowns list.If after that draft you also need to merge three PDFs, pull calendar items, and produce a formatted package with intermediate steps, then consider Work on a plan that allows it. The Chat draft still travels with you as the language spine.
Migration notes without the panic
People hit three confusing states in 2026 desktop landings:
- Old ChatGPT desktop (“Classic”) still installed while the new combined app is separate. Follow OpenAI’s migration prompt rather than running both forever.
- Codex app updates into ChatGPT desktop. That is intentional product merge behavior, not you installing the wrong thing.
- Mode labels lag the blog post. Community threads sometimes report missing Chat buttons or only Work + Codex on a build while web still shows Chat. Use web Chat if conversation is missing, and keep the three-job map in your head even when a button is temporarily absent.
None of that changes the teaching point: conversation, multi-step office agent work, and coding are different jobs. The UI is catching up to that split.
What you are not doing in this part
- Full Work approval loops and enterprise connector setup (later Work tutorial parts)
- Repo hygiene, tests, and PR review discipline (later Codex tutorial parts)
- Custom GPTs as a product (separate series)
- API keys and builder tools (product map later)
If you want the parallel story in Anthropic’s product family (chat vs Code vs Cowork), keep Learn Claude and the Claude product map nearby. Same judgment muscle, different labels and install paths.
Common mistakes on desktop week one
| Mistake | What happens | Fix |
|---|---|---|
| Treat every mode as “better chat” | Wrong tool, wasted usage | Use the three-job table |
| Start Work for a one-line rewrite | Overhead and burn | Stay in Chat |
| Start Codex for leadership prose | Repo-shaped friction | Chat (or Work if multi-step office) |
| Ignore Free limits | Mid-task walls feel like outages | Scope agent tasks; upgrade only when weekly work needs it |
| Grant every permission on day one | Risk without teaching value | Chat first; add grants when a real job needs them |
| Hardcode a model name in team docs | Docs rot when UI renames | Describe the job; note “current model string in UI” |
How to practice this week
- Install or open the official desktop app and list the modes you actually see.
- Run one pure Chat task (rewrite or explain) and export it the same way you did in Part 3.
- If your plan allows, open Work once with a tiny multi-step job (for example: “turn these three bullets into an outline + checklist, then stop”). Cancel if it starts wandering.
- If you write code for work, open Codex once against a safe toy repo or a throwaway folder. Do not point it at production secrets on day one.
- Write a personal sticky: “Chat for language. Work for multi-step office. Codex for code.” Stick it until the habit sticks.
FAQ
Do I need desktop if web Chat is enough?
No. Desktop is optional until Work or Codex is part of your weekly work, or you simply prefer a dedicated window. Part 3’s web and mobile habits still count.
Is Work the same as Chat with a different skin?
No. Chat is turn-by-turn conversation you drive. Work is built for multi-step agentic office tasks. You still review. The cost and failure modes differ.
Is Codex only for professional software engineers?
It is aimed at coding workflows. Analysts who live in scripts and repos can use it. People who only need prose should not force every task into Codex because the mode exists.
Why do people mention GPT-5.6 with Work and Codex?
OpenAI’s public product notes around the desktop merge and coding features have referenced that model family for some capabilities (for example computer use notes). Names change. Use the model label in your UI and OpenAI’s current docs when accuracy matters for a team standard.
Where is the Claude version of this map?
Learn Claude from scratch and the Claude product map series cover chat, Code, and Cowork-style work on Anthropic’s side. Same “pick the door for the job” lesson.
Quick recap
- Desktop ChatGPT (Mac/Windows) is where Chat, Work, and Codex live together; download from official OpenAI pages.
- Chat = conversation. Work = multi-step office agent work. Codex = coding.
- Free can open more doors than it can fully fuel. Usage burns faster on Work and Codex.
- Pick the mode from the job, not from which button looks newest.
- Hedge model names; trust the job map and the UI you see today.
Next: Part 5, Memory, custom instructions, and Projects, covers sticky context so you stop re-pasting house rules every morning. For Claude’s parallel track: Learn Claude. More paths: Learn hub.
Sources
Desktop packaging and plan limits change. Re-check official pages when you set team standards.
- OpenAI: ChatGPT for your most ambitious work (desktop app with Chat, Work, and Codex; Free availability and Codex app merge)
- OpenAI Help: Moving to the new ChatGPT desktop app (Classic vs new app; Codex migration notes)
- Download ChatGPT (desktop and mobile download hub)
- ChatGPT Learn: What’s new (desktop Codex merge notes and related feature blurbs)
- ChatGPT web (conversation baseline when desktop modes lag or differ)
- OpenAI Help Center (plans, usage, account help)
- Analytics Made Simple: Learn Claude from scratch (sibling product path)
- Analytics Made Simple: Practical AI (vendor-neutral verification habits)
- Analytics Made Simple: Learn (learning hub)
