The complete ChatGPT guide

Every plan, every model, every feature. Verified 22 September 2026.

Start here: it isn't a chatbot any more 🧭

If you last used ChatGPT as a text box that answers questions, the thing has changed shape underneath you. It's now three surfaces that share one subscription:

• Chat — fast, conversational, everyday questions. The thing you already know.
• ChatGPT Work — an agent built for long, multi-step jobs that end in a finished deliverable. A document, a spreadsheet, a deck, a research report.
• Codex — the coding agent. Writes, reviews, tests and ships code, with access to your actual files and terminal.

You switch between them inside the same app. Work and Codex aren't separate products you buy — they're modes, included across plans at different usage levels.

That's the single most useful thing to understand. Most people are still typing into Chat and wondering why it won't finish a real piece of work. It's because they're in the wrong mode.

Getting started on every platform 📲

• Web — chatgpt.com. Nothing to install. Everything works here except Codex-as-a-mode, the built-in browser and Computer Use.
• Desktop (Mac and Windows) — this is the one to install. The current desktop app bundles Chat, Work and Codex together, and it's the only place several features exist at all. Launched globally 9 July 2026. macOS needs 14 or later.
• Mobile (iOS and Android) — Chat, Work, Voice, images, Projects. No Codex mode, no Computer Use.
• Linux — a preview exists (Ubuntu, Debian, Fedora, Arch), but Computer Use isn't in it yet and Wayland support is experimental. Treat it as early.
• Inside Word, Excel, PowerPoint and Google Sheets — sidebar add-ins, on every plan including Free. More on these further down.

The plans, and what you actually get 💳

Prices verified 22 September 2026.

What each tier adds, in OpenAI's own bullets:

• Free — unlimited text chats on GPT-5.6 Luna · limited messages with uploads · limited, slower image generation · limited voice · limited deep research · limited memory · limited Codex
• Go — more messages with tools · more uploads · more image creation · more voice · longer memory
• Plus — advanced reasoning models · expanded messages and uploads · more complex image creation · expanded deep research · expanded memory · Projects, scheduled tasks and custom GPTs · expanded Codex · early access to new features
• Pro — 5× more usage · Pro reasoning · unlimited and faster image creation · maximum deep research · maximum memory · research previews
• Enterprise adds SCIM, EKM, domain verification, role-based access, data residency in ten regions, 24/7 support and SLAs

The models, explained properly 🧠

Two families are live.

• GPT-6 Astra — "our most capable model, built for the hardest end-to-end work." System card published 3 September 2026.
• GPT-6 Pro — Astra run at Pro effort, for difficult tasks and long-running workflows.

GPT-5.6 (published 9 July 2026)

• GPT-5.6 Sol — flagship for complex professional work: coding, research, science, design.
• GPT-5.6 Sol Pro — the highest-capability 5.6 option.
• GPT-5.6 Terra — the balanced everyday one.
• GPT-5.6 Luna — fastest and cheapest. This is what Free and Go run.

How you actually pick a model: the thinking slider 🎚

This trips up everyone coming back after a few months. On paid plans you mostly don't choose a model name any more. You choose how hard it should think, and ChatGPT routes you. The levels: Instant · Medium · High · Extra High · Pro

On Free and Go, "Think" runs GPT-5.6 Luna rather than Sol — so it's thinking, but with the small model.

The practical rule: Instant for anything you'd ask a colleague in passing. Medium or High the moment the answer has consequences. Most people leave it on Instant and then complain the output is shallow.

Context windows: how much it can hold 📏

This is the number that decides whether you can throw a whole contract at it.

Two things worth noticing. Go and Plus have identical context windows — you're paying the extra $12 for usage volume and model access, not memory. And Pro's reasoning window (400K) is bigger than Enterprise's (256K), despite Enterprise marketing promising an "expanded context window." Both numbers come straight off OpenAI's pages.

The API is a different animal: the flagship models report a ~1,050,000-token context window there. Don't mix these up — that's roughly 2.6× the largest window you can get inside ChatGPT itself.

Ads, and how to get rid of them 📢

Yes, ChatGPT has ads now. This surprises people, so here are the facts.

• Who sees them: "Ads may appear for users on the Free and Go plans. Plus, Pro, Business, Enterprise, and Edu accounts will not have ads."
• When it started: a US test on 9 February 2026. By 31 August 2026 it had reached 40+ countries.
• What they look like: "Ads can appear below the end of a response. Ads are clearly labeled as sponsored and visually separated from ChatGPT's response."
• What advertisers get: "We do not share your conversations with ChatGPT with advertisers, and we never sell your data to advertisers."
• Under-18s: no ads shown to accounts identified as under 18.

Three ways to reduce or remove them:

1. 1Settings → Ad Controls — toggles for "Personalize ads" and "Past chats and memory." Important: this turns off personalisation, not the ads. You'll still see ads, just less relevant ones.
2. 2Upgrade to Plus or Pro. Both are ad-free.
3. 3Stay on Free and take the trade. There's a no-ads option on the free tier that accepts "fewer messages per day and no access to tools like images." Shown in-product as "Reduce message limits." Almost nobody knows this exists.

ChatGPT Work 🛠

The mode most people haven't tried, and the one that changes what the subscription is worth.

Chat is "for fast, conversational assistance and everyday questions." Work is "an agent designed for longer, multi-step work and finished deliverables." It plans, works through steps, uses tools, and hands you a finished file rather than a wall of text.

It creates and edits documents, spreadsheets, presentations, reports and analyses — natively in Google Docs, Sheets and Slides, and in Excel via the desktop add-in. You can hand it a template as a .docx, .xlsx, .pptx or a Google file link and it'll work inside your format.

Availability: all plans on macOS and Windows desktop; Plus, Pro, Business, Enterprise and Edu on web and mobile.

Prompts worth stealing:

Read the three files in this project, then build me a one-pagesummary for a client who has ten minutes and no background.Plain language, no jargon, one clear recommendation at the end.Give it to me as a .docx.

Take this spreadsheet, find every row where the margin is under15%, and build me a second tab that groups those by supplier witha total per supplier and a short note on what's driving it.

Codex 💻

OpenAI's coding agent. "Codex is an AI agent that helps you write, review, and ship code." It reaches your local folders, repositories, terminals and developer tools.

It's on every plan, including Free — Free gets limited access, Plus expanded, Pro maximum.

Where it runs: the desktop app, an IDE extension, a CLI, the web, an SDK, and a Chrome extension. There's also Codex Remote — pair your phone to your Mac or Windows machine with a QR code and steer it from your pocket.

The desktop app 🖥

Worth installing even if you like the web version, because four things exist nowhere else:

• Codex as a mode
• The built-in browser
• Computer Use
• Site tools

The built-in browser 🌐

A real browser inside the desktop app. "Browse, sign in, download files, and review pages without leaving ChatGPT." You and ChatGPT look at the same page. It can work across tabs, download files, and wait while you sign in.

It uses its own browser state, not your Chrome profile — so it isn't quietly logged into everything you are. It asks permission before each new website. It supports sign-in, autofill, password management and extensions, and has an annotation mode for pointing at things on a page.

Cloud browser ☁️

Different thing. Where the built-in browser runs on your machine, the cloud browser gives ChatGPT "its own browser on a separate computer in the cloud." It reads pages, clicks buttons, fills forms and works through steps on public and signed-in sites — and "can continue in the background, including after you close your computer or turn off your phone."

Available in ChatGPT Work on paid plans, excluding Free and Go. You can't toggle it manually; it gets selected automatically when the job suits it.

It asks for confirmation before anything creating "a financial, legal, account, or other real-world commitment." And plenty of sites block automated browsers outright, so it won't work everywhere.

Computer Use 🖱

ChatGPT operating your actual computer. "ChatGPT can see and operate graphical user interfaces on macOS or Windows" — clicking, typing, moving files, changing settings, driving apps that have no API.

Where: in ChatGPT Work and Codex, with the Computer Use plugin installed. macOS and Windows only — not in the Linux preview.

Setup: macOS needs Screen Recording and Accessibility permissions.

• On Windows it "runs on the active desktop" — it can't work in the background while you use the same session. You hand over the machine.
• macOS supports locked computer use, so it can keep going after your Mac locks. Windows can't.
• It can't automate terminal apps or ChatGPT itself, can't authenticate as administrator, and can't approve system permission prompts.
• Admins can disable it or allowlist specific apps.

Deep research 🔎

"Plan, research, and synthesize complex questions into a documented report." You describe what you need, choose sources, review the plan it proposes, then get a structured report with citations.

Every tier has it, scaled: Free limited → Plus expanded → Pro maximum.

Voice 🎙

The feature is ChatGPT Voice. It has three options: Live, Advanced and Standard.

Live is the current one, powered by the GPT-Live-1 model family (announced 8 July 2026). It's full-duplex — "both models can listen and speak at the same time." It decides many times a second whether to speak, keep listening, pause, interrupt or call a tool, and it can hand a hard question to a bigger model mid-conversation. It handles text and images in the same chat, uses web search and memory, and Files and Projects work in it.

Images 🎨

Current version: ChatGPT Images 2.5, released 8 September 2026. It replaced ChatGPT Images 2.0 (21 April 2026).

"Sharper details, more precise editing, and faster generation" — up to 50% faster than 2.0, and better at preserving subjects in your reference photos. Under the hood the API models are GPT-Image-2.5 Flare (default) and GPT-Image-2.5 Sunburst (premium, tighter editing control).

What it does: text-to-image, editing an existing image, adding text, transparent backgrounds, templates for posters and logos, aspect-ratio control, sketch-to-image on mobile, and mobile editing with comments.

On every tier, scaled: Free limited and slower → Go more → Plus more complex and accurate → Pro unlimited and faster. Images with thinking — where it plans and refines before generating — is on Plus, Pro and Business.

Sites 🌍

ChatGPT Sites, currently in public beta. "Create, preview, publish, and share interactive websites and lightweight apps" — dashboards, project trackers, launch calendars, prototypes, internal portals, reports. No code. OpenAI hosts it.

Plus, Pro and workspaces only — not Free or Go. And not available in the EEA, Switzerland or the UK at launch.

Usage is "included up to plan-specific limits" which OpenAI hasn't published during beta. Custom domains work, but "Sites does not register a domain for you. You must already own the domain" — and custom domains aren't available in Enterprise workspaces yet.

Plugins, apps and skills 🔌

The naming here has changed twice in a year, so let's be precise.

Plugins is the umbrella term. "Plugins help ChatGPT and Codex complete workflows by packaging reusable instructions, connected tools, and other capabilities."

A plugin can contain three kinds of thing:

• Skills — instructions and workflow guidance
• Connected apps — link an external account so ChatGPT can read from it or act in it
• App templates — for admins configuring apps across a workspace

OpenAI's own distinction: "An app connects ChatGPT to an external service, such as Google Drive or Slack... A plugin packages capabilities into a workflow and can include one or more apps, skills, or both." Some plugins are skills only and need no account connection at all.

There are 1,400+ plugins in the directory. Since 17 September 2026 you can connect a personal and a work account to the same plugin.

Custom GPTs are being retired ⏳

• 26 October 2026 — creation of new custom GPTs ends
• 11 December 2026 — custom GPTs retire entirely

At retirement they "stop running and leave the GPT directory." OpenAI shipped migration tooling on 22 September 2026 to move them into plugins; migrated originals become read-only.

If a custom GPT is load-bearing in your workflow, migrate it now rather than in December.

Memory and personalisation 🧩

Memory draws on "past chats, saved memories, custom instructions, Library files, and connected apps like Gmail."

It isn't total recall: "Memory does not retain every detail from every conversation. ChatGPT decides which available information is relevant to a response."

Three parts: saved memories (explicit or auto-saved), chat history references, and a memory summary you can read and edit in Settings → Personalization → Memory.

Your controls — more than most people realise:

Projects 📁

The most underused feature on this page, and my pick for the one that'll change your week.

"Projects keep related chats, files, and instructions together." Three things make them work:

1. 1Project instructions. "Project instructions apply only within that project and override your global custom instructions." Set the rules once for this body of work.
2. 2Files. Upload PDFs, spreadsheets, docs, images, or paste text as reference material every chat in that project can see.
3. 3Project-only memory. "Your previously saved memories are not referenced in project chats" and "chats cannot reference conversations outside the project." Clean room. Since 14 August 2026 you can flip an existing project between default and project-only memory.

Library 📚

Not just images, despite what people assume. "Library lets you find and reuse files that you upload to or create in ChatGPT" — documents, spreadsheets, presentations, PDFs and images, with generated images getting their own section.

It can also surface files from connected Google Drive, Box, Dropbox and SharePoint accounts. Sharing arrived 9 September 2026 with Viewer and Editor access levels; shared items land in "Shared with me."

Scheduled tasks ⏰

ChatGPT running things without you. "ChatGPT can run one-time or recurring tasks, monitor for changes, and respond to supported events."

Data analysis, files and the Office add-ins 📊

Data analysis. "ChatGPT can run code in a secure environment to analyze and visualize data from spreadsheets, CSVs, and other structured formats." Interactive charts — bar, line, pie, scatter — arrived 8 June 2026.

The Office add-ins — all free, all overlooked:

Business and Enterprise 🏢

ChatGPT Business — $20/user/month billed annually, $25 monthly. Premium seats $100/$125. Minimum two paid seats, maximum 200 for workspaces created after 24 August 2026.

Premium seats (live 25 August 2026) give 5× Standard usage with no five-hour limit, weekly resets, and shared workspace credits.

ChatGPT Enterprise — custom pricing, no published figure and no published seat minimum.

Privacy, data and safety 🔒

The single most important setting. On personal accounts, your conversations are used to improve OpenAI's models unless you turn that off. The setting is called "Improve the model for everyone", and it lives in Settings → Data Controls.

Turning it off doesn't hide your history — chats "still show up in your history, but they won't be used to improve ChatGPT."

Temporary chat now has two modes — Personalized and Unpersonalized, chosen before your first message. A temporary chat "stays out of your chat history and is not used to improve OpenAI models while it remains temporary."

Teens and age. OpenAI runs age prediction — it "may use signals to predict whether an account belongs to someone under 18", including topics discussed, times of day and account age, independent of the birth date at signup. Accounts predicted under 18 get content protections and see no ads. Adults wrongly flagged can verify through a third party.

Parental controls link a parent and teen account from Settings → Parental controls → Add family member. Controls cover sensitive content, model training, memory, Codex network access, cloud browser, voice, image generation, study mode, and quiet hours that block access entirely during set times.

Prompting in the reasoning era ✍️

Prompting advice from 2023 is actively counterproductive now. The models reason. You no longer need to trick them.

Stop doing this:

• "Act as a world-class expert in…" — it doesn't help any more
• "Take a deep breath and think step by step" — it thinks anyway, on the levels that think
• Stacking five adjectives about tone
• Splitting one task into ten tiny prompts

Do this instead. Give it the four things it can't infer:

CONTEXT: who this is for and what's already been decided.TASK: the one thing you want, stated plainly.CONSTRAINTS: length, format, what to avoid, what must stay true.DONE LOOKS LIKE: what a good answer contains.

Compare:

Write a professional email to a client about the delay.

Our delivery slips from the 14th to the 28th because our suppliermissed a shipment. The client is a long-term one who's been patientbefore and will mostly care about whether the new date is real.Write the email: under 150 words, no apologising twice, give thenew date in the first two lines, and offer one concrete thing thatmakes the wait easier.

25 things worth knowing 🔑

1. 1Put standing instructions in a Project, not in every chat.
2. 2Use the thinking slider. Instant is the default and it's the wrong setting for real work.
3. 3Automatic thinking-upgrade is gone. Nothing raises the effort for you any more.
4. 4Custom instructions allow 5,000 characters. Most people use 200.
5. 5Project instructions override global ones — you can have different rules per body of work.
6. 6Project-only memory keeps client work sealed off from everything else.
7. 7Temporary chat ≠ private. Retained up to 30 days.
8. 8"Improve the model for everyone" is on until you turn it off.
9. 9On Free, you can kill ads by accepting lower limits — "Reduce message limits."
10. 10Ad Controls only turns off personalisation, not the ads themselves.
11. 11Go and Plus have identical context windows. You're buying usage, not memory.
12. 12Pro's reasoning context (400K) beats Enterprise's (256K).
13. 13Review the deep research plan before it runs.
14. 14Never type a password into the chat — only into the built-in browser.
15. 15The cloud browser keeps working after you close your laptop.
16. 16On Windows, Computer Use takes over your desktop. Plan around it.
17. 17The Office add-ins are free on every plan, including Free.
18. 18ChatGPT for Word turns itself on 1 October 2026.
19. 19Spreadsheet chats don't appear in your main history.
20. 20Library holds documents, not just images — and Pro gets 100 GB.
21. 21Ask "what do you remember about me?" then prune it.
22. 22Say "don't mention this again" to kill a topic permanently.
23. 23Scheduled tasks can trigger on Gmail, Slack and GitHub — not just clocks.
24. 24Migrate custom GPTs before 11 December 2026.
25. 25Move off GPT-5.5 before 14 October 2026.

Your first week 📅

If you're starting from zero, do these in order. One a day.

• Day 1 — Set it up properly. Install the desktop app. Write your custom instructions — who you are, what you work on, how you want to be spoken to. Use more than two lines.
• Day 2 — Turn the thinking up. Take a question you asked on Instant and ask it again on High. Notice the difference. That's the gap most people never see.
• Day 3 — Make a Project. Pick something ongoing. Add instructions and two or three reference files. Work inside it for the rest of the week.
• Day 4 — Try Work, not Chat. Give it a real multi-step job that ends in a file. A summary, an analysis, a deck.
• Day 5 — Check your privacy. Settings → Data Controls. Decide about "Improve the model for everyone." Read your memory summary and delete what's wrong.
• Day 6 — Automate one thing. A scheduled task. A weekly summary, a morning digest, a recurring check.
• Day 7 — Find its edge. Give it something you already know the answer to, and see where it goes wrong. That's the most valuable hour you'll spend, because it calibrates everything after.

The honest bit 🙏

Sources 🔗

Everything above comes from OpenAI's own pages — no third-party blogs were used for any fact.