ChatGPT Complete Guide: Plans, Models, Work, Codex, and Privacy Settings

What This Document Is: What ChatGPT Looks Like Now

This is a continuously updated reference document, not a one-time tutorial. The content was queried on 2026-11-06, and every fact is based on OpenAI’s official help pages, official documentation site, and API documentation. Where I could not find something, or where official pages contradict each other, I say “not published by OpenAI” or note the inconsistency directly, and I do not fill the gap with third-party numbers.

One premise first: ChatGPT changed enormously this past year. If your picture of it is still GPT-4o, o3, Canvas, Agent mode, and GPTs, most of that no longer matches. Quotas, model names, and regional restrictions get adjusted within a month or two, so treat the numbers in this guide as a snapshot “as of the query date”, and check the official pages again before you act.

The three modes of ChatGPT

OpenAI now defines ChatGPT as an AI agent you operate in natural language, split into three modes:

In the desktop app, Work and Codex overlap: Work hides most technical details and emphasizes the deliverable, while Codex shows diffs, the review view, and the Pull requests panel.

The new ChatGPT desktop app (macOS and Windows, with documentation for a Linux version as well) includes all three modes: Chat, Work, and Codex. The original Codex app became this new app after an update. The old ChatGPT desktop app was renamed ChatGPT Classic; it still receives model updates and security patches, but new agent features may appear only in the new app.

Work runs in the cloud on web and mobile. In the desktop app, when your plan and workspace allow it, Work can use local files and desktop apps with your authorization. Codex is a separate view on desktop, and its history is kept apart from ChatGPT history. You cannot select Codex on web or mobile, but you can reach Codex conversations on your desktop through the Remote tab of the mobile app, and you can also start Codex Cloud tasks.

What changed this year

As of the query date, here is how the official pages differ from the 2025 picture:

How to read this guide

Suggested order:

1. Chapters 2 and 3 first, to work out which models your plan can use and how much quota you get.
2. Chapters 4 to 6 cover the features you use in everyday conversation, voice, and personalization and memory.
3. Chapters 7 to 9 cover long tasks: Deep research, Work, Plugins and MCP, Dots, Space, and Sites.
4. Chapter 10 is for developers: the entry points for Codex and the API.
5. Chapter 11 covers privacy and data controls, and I recommend everyone read it.
6. Chapters 12 and 13 cover prompting tips, frequently asked questions, and how ChatGPT is positioned differently from Claude and Gemini.

Every chapter ends with its sources and the query date. If a number or model name differs from what you see in the interface, trust the interface and the official pages. That is exactly why this guide needs to keep being updated.

A note on prices: USD prices come from official help pages. NT$ prices are the reference prices that chatgpt.com/pricing shows for Taiwan, and they are affected by region, IP, and in-app purchases through the App Store and Google Play. See Chapter 2 for details.

Source: ChatGPT release notes, Moving to the new ChatGPT desktop app, ChatGPT Work and Codex, Model Release Notes, ChatGPT agent, Creating and editing GPTs, Plugins in ChatGPT, ChatGPT Atlas retirement, Sora discontinuation, Retiring group chats, Use ChatGPT (queried 2026-11-06)

Plans and Pricing

ChatGPT plans are now split very finely. First confirm which tier you are on, so the models, features, and quotas in the rest of the guide line up.

USD pricing

Go, Plus, and Pro do not offer annual billing, and you cannot prepay several months at once.

Reference prices shown in Taiwan

The following are the prices chatgpt.com/pricing showed for Taiwan on 2026-11-06. They are reference prices only, and the actual amount is whatever you see at checkout. OpenAI also states that in-app subscriptions are managed by Apple or Google, and prices for store purchases may differ.

Changes to the three Pro tiers

Pro quotas have been adjusted several times in the past two months. This is the state as of the query date:

• Pro 200 once paused new subscriptions. The official release notes on 2026-09-29 say it has reopened, and usage changes are announced by email.
• On 2026-10-29, the Pro 200 quota reduction was reversed: people who already had the higher quota keep it, and those on the lower one are doubled to catch up. The Work and Codex quotas for Pro 100 and Pro 500 were also doubled.
• Pro currently has no 5-hour limit in Work and Codex, but there may be a weekly limit.

Features and context by plan

OpenAI’s wording is “Unlimited text chats … subject to abuse guardrails”: only Instant text chat is unlimited, while reasoning, Pro, Work, Codex, images, voice, and uploads each have their own limits. Free’s Instant context is only 27K, so check your plan before you process long documents.

How usage works

• Credits: Plus and Pro can buy credits to keep going. The flow is that you use up the included quota first, and then credits are deducted. Credits apply to Codex, ChatGPT Work, and ChatGPT for Word / Excel / PowerPoint, and not every model feature can use them.
• Speed modes: Fast consumes included quota at 2.5 times the rate. Ultrafast (only Astra and GPT-6.1 Sol, and only Pro 500 and eligible Enterprise / Edu) consumes included quota at 8 times the rate, and 6 times the credit rate when you pay with credits.
• Message estimates for Work and Codex: For Plus and Business Standard, per 5 hours that is roughly 5 to 45 messages for Astra, 15 to 160 for GPT-6.1 Sol, 35 to 285 for GPT-6.1 Terra, 15 to 150 for GPT-6 Sol, and 350 to 3,000 for GPT-6 Luna. These are estimates, not fixed caps, and there may also be weekly limits.
• Images and voice: Image generation uses quota about 3 to 5 times faster than a normal turn. Desktop voice costs about $0.05 / minute.
• Shared quota: OpenAI says that from 2026-12, usage quotas for Plus, Pro, and Business will become shared between ChatGPT and Codex and counted in tokens. Instant chat does not count against quota. You can see combined usage and reset times in Settings → Usage. OpenAI has not yet published the conversion details, so this is the part most likely to change after the guide is published.
• When you hit a limit, you can only wait for the reset, switch to another model, upgrade, or buy credits. There is no setting that raises the limit, and Support cannot reset your usage.

What Free and Go actually include

• Free: Search, data analysis, file and image uploads, image generation, 500 MB of Library storage, and access to Think (already in the mobile app, rolling out to web).
• Go: Everything in Free, plus higher limits for images, file uploads, data analysis, and voice, and longer memory. It has projects, tasks, and Library (limited by storage). You can use Think, but you cannot use legacy models. The voice quota is the same as Free.
• Ads: Free and Go may show ads. Plus, Pro, Business, Enterprise, and Edu have no ads. Temporary Chat shows no ads.

Business and Enterprise

• Business: For teams of 2 to 200 people. Standard and Premium seats can be mixed. Premium has no 5-hour limit and includes dots. It includes SAML SSO, MFA, no training on your data by default, ChatGPT Space, usage analytics, and spend controls. Pro messages in Chat: Standard gets 15 per month and Premium gets 50 per week, shared between GPT-6 Pro and GPT-5.6 Sol Pro.
• Enterprise: SCIM, EKM, RBAC, the Compliance API, data residency (the official page says ten regions), custom retention policies, and an SLA.
• Nonprofit organizations get up to 75% off Business and Enterprise, and certified US K-12 educators get it free until 2027-06.

Source: ChatGPT pricing (Individual and Business & Enterprise tabs), About ChatGPT Pro tiers, What is ChatGPT Plus?, What is ChatGPT Go?, ChatGPT Free Tier FAQ, Understanding ChatGPT and Codex usage limits, ChatGPT Business models and limits, Codex pricing, the 2026-09-29 and 2026-10-29 entries in the ChatGPT release notes, Ads in ChatGPT (queried 2026-11-06)

Models: the GPT-6 Family and the Thinking Slider

ChatGPT’s models fall on two lines: the Chat model menu, and the model picker in Work and Codex. The two sides use different models, and this is the easiest place to get confused.

Models in Chat

From 2026-10-07, GPT-6 in the model menu replaced the old Latest. An old Latest selection moves to GPT-6 at the same effort. People who explicitly chose GPT-5.6 Sol stay as they are. Older selectable models are not retired just because GPT-6 launched.

After you pick GPT-6, use the Thinking slider to adjust how much it thinks:

Which model sits underneath: on paid plans, Instant through Extra High are powered by GPT-6 Sol, but the interface shows only “GPT-6” and does not label Sol. The Pro position is GPT-6 Astra. GPT-6 on Free and Go is at the Luna level. Official pages label the version inconsistently (some say GPT-6 Luna, others say GPT-5.6 Luna), so this guide only says “Luna level”.

What each plan can choose

A few points that are easy to misread:

• GPT-6 Pro has a usage cap. A Pro subscription does not mean unlimited Pro messages. When Pro 200 reaches its weekly cap it automatically falls back to GPT-5.6 Thinking Medium. On Pro 100, GPT-6 Pro and GPT-5.6 Sol Pro share one weekly quota.
• From 2026-09-14, Plus and Pro no longer automatically switch from Instant to Thinking, and the “Higher intelligence” setting on web was removed with it (Business still keeps this setting in Settings → General).
• On 2026-10-29, GPT-6 for Free, Go, and logged-out users was changed to give more concise replies and be less sycophantic. The main model for Plus and Pro is unchanged.
• The 2026-10-16 update made GPT-6 Pro faster and better at using plugins and tools, and lets it open multiple agents to handle subtasks in parallel. It is being rolled out gradually to eligible Pro users.
• For the exact message cap of Medium and High in Chat on Plus, the official pages do not publish a number. They only say it depends on the plan, you may be downgraded to another model, and you can check Settings → Usage.
• If you cannot see GPT-6, that is because of the gradual rollout, and managed workspaces are also limited by model permissions. If you cannot see Astra, use Menu → Check for Updates in the desktop app, and the Codex CLI needs version 0.153.0 or later.
• Some high-risk biology and cybersecurity requests may be refused or need extra checks, and occasionally this also affects legitimate requests.

Models in Work and Codex

These models do not appear in the regular chat menu:

The Work picker defaults to the Power slider (Smarter to Faster), with six presets: Luna High, GPT-6 Sol Light, GPT-6 Sol Medium, Astra Light, Astra Medium, and Astra Extra High (some paid plans do not have Astra Extra High). Click “Advanced” to specify the model, reasoning effort, and speed yourself.

Reasoning effort

The effort levels in Work and Codex are Light (called Low in the CLI), Medium, High, Extra High, Max (gives a single task more reasoning time), and Ultra (processes in parallel with subagents). Luna goes up to Max at most. To see Ultra on the desktop app slider, turn it on in Settings → Configuration.

OpenAI’s selection suggestions, from small to large:

1. Luna Low: Fine edits, small-scope questions, simple data extraction
2. Luna Extra High: Finding the latest context across several apps, setting priorities
3. GPT-6.1 Sol Medium: Complex technical work, integrated deliverables you expect to revise again
4. GPT-6.1 Sol Extra High: Polishing a deliverable, making judgments under conflicting evidence
5. Astra Low: Concise writing and rewriting that keeps the facts intact
6. Astra Medium: Large projects that need broad context and reliable interaction
7. Astra Extra High: Demanding analysis and complex deliverables

OpenAI’s recommended method is to compare using the same input, fix the model and adjust only the effort, and keep the lightest setting that meets your bar. This is the same thinking as the chapter “Models, Effort Levels, and Fast Mode” in the Claude Code guide: do not default to the highest setting.

Models at the API level (brief)

Input over 272K tokens uses the long-context rate (for example Astra input $20, output $75). Batch and Flex cost about half the standard price. Regional processing and FedRAMP endpoints add 10%. Astra and Sol do not support none for reasoning effort (low to max), while Terra and Luna support none. The API and a ChatGPT subscription are billed separately.

Retired models

As of the query date, within a single month GPT-6.1 Sol and GPT-6.1 Terra went live and GPT-5.5 was retired. The mapping of model names will go stale quickly, so treat this chapter as a map for understanding the structure, not a permanent lookup table.

Source: GPT-6 and other models in ChatGPT, Model Release Notes, ChatGPT release notes, Retiring GPT-4o and other ChatGPT models, ChatGPT Business models and limits, Codex Models, Model selection, API Models, API Pricing, Using the latest model (queried 2026-11-06)

Everyday Chat Features: Search, Files, Data Analysis, and Images

This chapter covers the features you use most in Chat mode. Availability of each feature varies by plan and region, so rely on the official pages for details.

Web search

Free, Go, Plus, Pro, Business, Enterprise, and Edu all have search, and it works even when you are not signed in, on web, desktop, and mobile apps. ChatGPT decides automatically whether to search. You can also pick Search manually with /, or open it from the tools menu via View all tools → Search.

• Replies include citations, and you can click Sources to see all of them. OpenAI’s reminder: search results and citations may be incomplete, outdated, or wrong, so open the sources and check.
• Your query is rewritten and then sent to search partners (the page lists the privacy statements of Microsoft and Shopify).
• Approximate location comes from your IP, and precise device location is off by default (Settings → Data controls → Location). If Memory is on, memory may be used to rewrite the query.
• Restaurant reservations (OpenTable worldwide, Resy in the US, Yelp in the US and Canada) are limited to consumer plans and do not apply to Work.

File and audio uploads

Limits:

• Documents and slides are 512 MB per file. Text and documents are 2M tokens per file. CSV and spreadsheets are about 50 MB. Images are 20 MB. Audio is 512 MB per file.
• At most 80 files per 3 hours, and Free gets 3 per day.
• Storage cap is 25 GB per user and 100 GB per organization.
• Audio upload (launched 2026-10-06) is for paid plans only, and Free does not have it.
• File count in Projects: Free 5, Go and Plus 25, Edu, Pro, Business, and Enterprise 40.

OpenAI warns that transcription can be wrong, speaker identification is unreliable, and performance varies across languages, so check important content yourself.

Data analysis

It can inspect uploaded data and produce tables and charts. This guide did not check the details on OpenAI’s help page (supported file sizes, chart types, and so on) paragraph by paragraph, so refer to the official Data analysis with ChatGPT page before you use it.

Image generation

The current version is ChatGPT Images 2.6 (2026-11-04): faster, and it preserves more detail across edits. It adds GIF creation and saved reference images, which you can cite later in a conversation with @name.

• Every plan can use it. “Images with thinking” is offered on Plus, Pro, and Business, and OpenAI says it will come to Enterprise and Edu later.
• Features include selecting an area to edit, any aspect ratio, templates (Images → Templates), sketching on your phone to generate an image, adding text, and transparent backgrounds.

Writing blocks and code blocks

After Canvas was retired, it was replaced by blocks inside the chat (from 2026-05-28):

• Writing block: First drafts of emails, posts, and documents can be edited right in the conversation, with full screen, undo, and redo. Some can be sent as email or saved to Library directly.
• Code block: Can be copied, edited, and shown full screen. Preview supports HTML, React, SVG, Mermaid, Vega, and Vega-Lite. Python supports Run (sandboxed execution).
• Edits inside Temporary Chat may not be kept.

Paid users can keep using canvas for now through legacy models, until those models are removed. Some official pages (Projects, GPTs) still mention Canvas, and the Model Release Notes of 2026-05-28 are the authority.

Intelligent UI

Exclusive to GPT-6: answers can include text, visuals, and interactive components, such as side-by-side comparisons, charts, calculators, bill-splitting tools, and small games, and they can start displaying while the reply is still thinking.

• It does not support the Pro position, the Work tab, or Voice.
• On web you can choose Simple in Settings → Personalization → Layout and visuals to reduce visual elements.
• It has no separate quota. Old macOS and Windows desktop apps do not support it, so use web or update the app.

Other features

These features are recorded in the release notes, and this guide lists only their names and dates: Flashcards (2026-09-22, all plans), Study mode, Health and Finances (mainly in the US), Shopping try-on, Locked chats (2026-10-20, see Chapter 11), account switching (2026-10-30, iOS and Android, up to 2 accounts), and ChatGPT for Word / Excel / PowerPoint (Microsoft add-ins; Word went live on 2026-09-17, all plans including Free).

Source: Searching the web with ChatGPT, Uploading files and audio to ChatGPT, Data analysis with ChatGPT, Images in ChatGPT, Working with writing blocks and code blocks, Intelligent UI in ChatGPT, Model Release Notes, ChatGPT release notes (queried 2026-11-06)

Voice: ChatGPT Voice

The entry point for voice is Settings → Voice. ChatGPT Voice now has three modes, with different purposes and different quotas.

The three voice modes

On 2026-10-15, new Live voices were added, including French, German, Hindi, Japanese, Korean, Mexican Spanish, and the English voice Bubbie. The options shown can differ by region. Live began supporting plugins on 2026-09-23.

Voice hours by plan

Chat voice has a rolling quota over each 24 hours:

Business uses credits after the quota is exceeded, at a rate of 1.25 credits / minute. This guide did not obtain the voice quota for Pro 500 from the official pages.

Using voice in Work

You can use voice inside Work (web, mobile, desktop). Actions that need approval are confirmed on screen, and spoken approval is not supported. Intelligent UI does not support Voice (see Chapter 4).

Other voice entry points

OpenAI also lists a 1-800-ChatGPT phone and messaging entry point, and ChatGPT Dictation (recording to text). For these two, this guide only confirmed that they are listed in the help center. See the official pages for the conditions of use.

Source: ChatGPT Voice, the 2026-09-23 and 2026-10-15 entries in the ChatGPT release notes, Help center ChatGPT collection (queried 2026-11-06)

Personalization and Memory: Custom Instructions, Memory, Projects

There are four layers for making ChatGPT understand you better the more you use it: Custom Instructions (global instructions), Memory, Projects, and Library and Space (file spaces). Project instructions override the global custom instructions, and knowing where each layer’s boundary lies works better than re-explaining your background over and over.

Custom Instructions

• Available on all plans and all platforms. Once saved, they apply immediately to all chats, including existing conversations.
• Character limits: Free and Go get 1,500 characters. Plus, Pro, Enterprise, Business, and Edu get 5,000 characters.
• Path: Settings → Personalization (Customize ChatGPT in the mobile app).
• They are not shared with viewers of a shared link.
• The API has no custom instructions, so you put them in the system message yourself.
• OpenAI’s warning: if you use third-party plug-ins, the model may pass related information from your instructions to the plug-in developer, so do not put sensitive information in them.

Personality and Characteristics

Settings → Personalization → Base style and tone lets you pick a base tone. There are also Characteristics, which adjust traits such as warmth and enthusiasm.

Memory

ChatGPT’s memory can come from past chats, saved memories, custom instructions, files in Library and Space, and connected apps (such as Gmail).

• Controls: Settings → Personalization → Memory. Depending on your plan and experience, you may see switches such as Memory, Reference saved memories, Reference chat history, Memory summary, Manage, and Saved memories. OpenAI says the details of each item vary by plan, region, platform, and workspace, and there is no per-plan feature table.
• Two mechanisms coexist: Improved Memory (a continuously updated summary) and legacy saved memories (explicit entries), and you can switch between them.
• Sources shows which personalization information a reply drew on.
• Project-only memory can restrict the memory context to a single project.
• Multiple accounts (2026-10-30): each account keeps its own chats, memory, settings, and subscription.

The trap in deleting memory

Deleting a chat does not automatically delete a standalone saved memory. To remove a piece of information completely, you need to deal with all of these: the memory summary, saved memories, regular and archived chats, files in Library and Space, and connected apps.

• After you turn off Reference chat history, memories derived from chat history are scheduled to be deleted from the system within 30 days.
• The log of deleted saved memories is kept for up to 30 days.
• Turning off Memory does not disable the safety uses in a very small number of high-risk situations.
• Chat history itself has no separate storage limit.

Projects

Projects keep related chats, files, and instructions together.

• Project instructions apply only inside that project, and they override the global custom instructions.
• You can add PDFs, spreadsheets, documents, and images, or paste text directly. You can paste Google Drive (file or folder) and Slack (channel) links as sources.
• Replies can be added to sources with “Add to project sources”. Existing chats can be dragged into a project (chats created with a GPT cannot be moved).
• Sharing: Free, Go, Plus, Pro, Business, Enterprise, and Edu all have project sharing, with Chat and Edit permissions.
• See Chapter 4 for file count limits.
• Scheduled tasks created inside a project cannot read the project’s files.

Library and Space

• Library: Stores files, saved products, and places. It can connect to Google Drive, Box, Dropbox, and SharePoint, and it can be shared (Viewer or Editor).
• Space replaces Library for accounts with access: a collection of Pages, files, slides, spreadsheets, and Sites, where you can invite collaborators. Existing Projects remain separate. As of the query date, Space is listed as a feature for Pro, Business, and Enterprise. See Chapter 9 for details.

Skills

Skills are reusable instructions that can be carried inside a plugin (Chapter 8 covers the structure of plugins again). OpenAI has a separate Skills in ChatGPT help page. This guide did not check its details paragraph by paragraph, so refer to that page before you use it.

Source: ChatGPT Custom Instructions, Memory in ChatGPT, Projects in ChatGPT, Using Library, Skills in ChatGPT, Scheduled tasks in ChatGPT, Personalize, Space, the 2026-09-29 and 2026-10-30 entries in the ChatGPT release notes (queried 2026-11-06)

Deep Research and ChatGPT Work

Search is fast and gives a short summary. For a more complete result there are two paths: Deep research produces a report with citations, and ChatGPT Work produces a deliverable. The latter is the successor to the old Agent mode.

Deep research

The flow: describe what you need, optionally choose specific sites, upload files, or connect apps, and ChatGPT first proposes a research plan (which you can edit). During the research you can watch progress in real time, interrupt midway and adjust sources, and in the end you get a structured report with citations.

• How to start: Type /Deepresearch, pick Deep research from the tools menu (+), or use the sidebar menu. In Work, open it from the + menu or Plugins.
• Restricting sites: Sites → Manage sites lets you limit it to specified sites, or choose “Prioritize these sites”, which favors those sites but still allows the whole web.
• Report: Shown full screen, with a table of contents, sources, and an activity log. It can be downloaded as Markdown, Word, or PDF.
• Connected apps: Only read actions are used, with no writes.
• You can choose a legacy model. Usage depends on your plan, and for plans with a fixed monthly quota it resets every 30 days (counted from first use). It is not available in every country.

ChatGPT Work

Work is the mode for long, multi-step tasks that produce deliverables (documents, spreadsheets, slides, reports, Sites).

• Execution environment: Cloud on web and mobile. On desktop you can choose cloud or local. Cloud Work conversations sync across web, mobile, and desktop.
• Cloud browser: Work has its own cloud browser that can read web pages, click, fill forms, and operate on sites you are logged in to. It pauses when it hits a login, an input, or something that needs confirmation. Available on paid plans only; Free and Go do not include it.
• Desktop features: Built-in browser, browser extensions (Chrome, Edge, Brave, Opera, Vivaldi), Computer Use, appshots, and site tools (WebMCP).
• Long tasks: On desktop, CLI, and IDE, you can use /goal to set a goal and verification conditions, and /plan to clarify first. On web, use Work and state the outcome, constraints, and review criteria in the prompt.

Do not write about the old agent mode as a current feature: the opening of OpenAI’s ChatGPT agent page already says agent is no longer available.

Safety: prompt injection

OpenAI’s safety advice for agent-type features was originally written on the ChatGPT agent page, and the principles apply equally to Work’s cloud browser:

• Enable only the apps the task needs.
• Avoid vague open-ended instructions, such as “check my email and handle everything”.
• Stop immediately when you notice suspicious behavior.
• For sensitive logins, use takeover and do them yourself.

Scheduled tasks

The entry point is Scheduled. You can create one-time, recurring, change-monitoring, and event-triggered tasks, and notifications can be Push or Email.

• Free can run at most once a day, and in a flexible time window (morning, afternoon, evening). Hourly and exact times need a paid plan.
• Event triggers (launched 2026-08-25): a new Gmail message, a new message in a Slack channel (you must add @ChatGPT to that channel), and GitHub PR activity. Available from Plus up; Free and Go do not have them.
• You can share a task link, and the other person creates an independent copy.
• Voice, GPTs, and Locked chats are not supported.
• Completed scheduled tasks can still consume quota in Chat.

OpenAI’s advice: for recurring tasks, tune the prompt in a normal chat first, and then schedule it.

How to choose

Source: Deep research in ChatGPT, ChatGPT Work and Codex, Using cloud browser in ChatGPT, ChatGPT agent, Scheduled tasks in ChatGPT, Features, Long-running work, Use ChatGPT (queried 2026-11-06)

Plugins, Apps, and MCP

What used to be connectors, apps, and GPTs has now been merged into plugins. This chapter covers the structure of a plugin, how to install and call one, custom MCP, and the retirement and migration of GPTs.

What a plugin is

A plugin is made up of these parts:

Installing and calling

• Sidebar Plugins directory → Install plugin → follow the prompts to Connect.
• In a conversation, mention one with @, or call it from the + menu. Some plugins are available on desktop only.
• The directory is visible on every plan, but whether you can install and use a plugin depends on your plan, workspace, role, and region. A common restriction on Go is that some features need Plus or higher.
• From 2026-09-17, all plans can connect multiple accounts (for example personal and work) to a plugin, no longer limited to Gmail, Google Calendar, and Contacts.
• Plugin Creator: After installing it from the directory, use @plugin-creator to build your own plugin through conversation.

Custom MCP server

The web version of Business and Enterprise / Edu supports full MCP:

1. Go to Plugins → + → Add custom MCP server.
2. Enter a public HTTPS MCP endpoint, or use Secure MCP Tunnel.
3. Confirm the risk warning.

On the Business plan only admins and owners can create and publish. Whether personal plans (Free, Go, Plus, Pro) can add a custom MCP is not mentioned on the official pages this guide checked, so go by the Plugins interface.

The official Dev Day page also mentions a draft spec for MCP Events, meant to let a plugin trigger automation when a connected app has an event. It is currently a draft, so do not treat it as a live feature.

GPT retirement

• Personal accounts (Free, Go, Plus, Pro) can no longer create or publish new GPTs.
• Existing GPTs still work, and can be edited depending on permissions. The mobile app can only use them, not create them.
• For affected Enterprise workspaces, the end date for creating new GPTs was 2026-10-26.
• The standard retirement date is 2026-12-11. For FedRAMP and workspaces with an approved extension, it is 2027-02-11.
• After migration, the original GPT becomes read-only.

Things to note when migrating:

• Custom actions do not carry over, so switch to an app connector or a custom MCP.
• Model selection, conversation starters, and chat history are not carried over.
• Before retirement, a GPT is limited to 20 knowledge files at 512 MB each.

As of the query date, the standard retirement date is only a little over a month away. If you have workflows that depend on GPTs, I suggest taking inventory first by following OpenAI’s Migrate custom GPTs guide.

Source: Plugins in ChatGPT, Connecting custom MCP servers in ChatGPT, Creating and editing GPTs, Migrate custom GPTs, DevDay 2026, the 2026-09-11 and 2026-09-17 entries in the ChatGPT release notes (queried 2026-11-06)

Dots, Space, and Sites

These three features mainly target plans from Pro up, and they have not been live for long. There are quite a few restrictions on region, plan, and billing terms, so be sure to confirm your own eligibility before using them.

Dots

Dots are always-on agents: each has its own cloud computer and browser, and keeps working toward the goal you gave it between conversations. The underlying model is GPT-6 Astra, and from 2026-11-04 GPT-6.1 Terra is also one of the driving models.

Who can use it

• Pro 100, 200, and 500 (age 18 and over, excluding the EEA, the UK, and Switzerland) and Business Premium.
• Enterprise, Edu, and Healthcare have it on by default from 2026-11-18 (the billing start date), unless the workspace opts out.
• UAE-resident workspaces are not eligible, and mobile web is not supported.

Usage

• Conversations with a dot do not count against ChatGPT quota. Tasks a dot starts in Work or Codex count against the corresponding quota.
• The first month after launch is extended quota (the first month does not count against plan quota). OpenAI has not yet published the terms for each plan after that.

Limits

• Dots make mistakes, and stopping a task does not undo actions that were already completed.
• Only one personal computer can be connected at a time (the app must be open).
• The cloud browser and local browser sessions are separate, and some sites block the cloud browser.

Controls and entry points: Pause, Activity, Scheduled, Take over. From 2026-10-28, Pro users in the US can text a dot by iMessage or RCS (beta, through the third-party messaging provider Linq).

Space

Space is a collaboration space for Pages, files, and slides. Sharing a Page does not expose your private chats or memory. On 2026-10-16, ChatGPT Slides was added, which can export to PowerPoint, PDF, or Google Slides, and is being opened gradually to eligible Pro, Business, and Enterprise users.

For how Space replaces Library, see Chapter 6.

Sites

Sites lets you build, host, and share websites through conversation, and it can also host existing compatible projects.

• Plans include a monthly hosting allowance. Hosting is free until 2026-12-31, and metering begins on 2027-01-01.
• Each site has a D1 database with 10 GB of capacity, and raw TCP is not supported.
• From 2026-10-15, Plus and Pro no longer have a fixed cap on the number of Sites (usage and storage limits still apply).
• From 2026-09-03, you can share with specified people outside the workspace.

Checklist before use

1. Confirm whether your plan and region are eligible.
2. Think through which actions must not be executed automatically, and Take over midway when needed.
3. If you use Sites, put the fact that metering starts on 2027-01-01 into your budget.

Source: Dots, Space, Sites, the 2026-09-03, 2026-09-29, 2026-10-15, 2026-10-16, 2026-10-28, and 2026-11-04 entries in the ChatGPT release notes (queried 2026-11-06)

For Developers: Codex and the API

ChatGPT’s entry point for developers is Codex, which is the same type of product as Claude Code. The API is a separate path with its own billing.

Codex availability and entry points

Codex is included in all ChatGPT plans (including Free and Go, with different usage). Codex Cloud is limited to eligible Plus, all Pro, Business, Enterprise, Healthcare, and Education, and does not include Free or Go.

OpenAI lists six entry points: the ChatGPT desktop app (macOS, Windows, Linux), the ChatGPT mobile app, ChatGPT web, the Codex CLI, the Codex IDE extension, and Codex Cloud.

Codex CLI

# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh

# Windows PowerShell
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

# or npm / Homebrew
npm install -g @openai/codex
brew install --cask codex

Run codex in your project directory and choose Sign in with ChatGPT to log in.

The local config file is config.toml (for example model = "gpt-6.1-sol"). Custom instructions go in AGENTS.md. Web search is controlled with web_search = "cached" | "live" | "disabled". Codex’s local memory is off by default, and you can control a single conversation with /memories.

IDE and Cloud

• IDE: VS Code (including Insiders), Cursor, and Windsurf install the Codex extension (openai.chatgpt). Xcode and JetBrains use their own agent settings.
• Cloud: You create and publish a “reusable cloud environment” (which describes the development setup, and Codex checks and tests it). Tasks run in an isolated workspace and can continue while your computer sleeps. When they finish, you review the changes and test results, then commit or open a PR.
• GitHub: Comment @codex review on a PR. Code Review counts toward Code Review usage only when it runs through GitHub.
• Auto-review (automatic approval review): When you sign in with ChatGPT, review checks are free and do not count toward usage.

Codex usage and rates

See Chapter 2 for message estimates. Standard rates in credits (per 1M tokens):

Compared with Claude Code

Codex’s effort levels are Light (called Low in the CLI), Medium, High, Extra High, Max, and Ultra. Claude Code has low through max plus ultracode. The concepts are similar and the names differ. For project instruction files, Codex uses AGENTS.md and Claude Code uses CLAUDE.md.

API entry points

The API is billed separately from a ChatGPT subscription, and Plus does not include the API. OpenAI recommends using the Responses API.

pip install openai        # Python
npm install openai        # Node.js
export OPENAI_API_KEY="your_api_key"

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6-astra",
    input="Write a one-sentence bedtime story about a unicorn.",
)
print(response.output_text)

See Chapter 3 for model prices. When migrating from older models, note that when effort is not none, you need to remove parameters such as temperature and top_p. minimal maps to low. For tool calls, OpenAI recommends the Responses API. If the GPT-6 family keeps stopping to ask for approval, OpenAI suggests adding instructions along the lines of initiative / follow-through to the prompt, and checking instruction files such as AGENTS.md, because Astra is more sensitive to them.

Other entry points

• Sign in with ChatGPT: Eligible Plus and Pro users can authorize participating apps to use their own ChatGPT quota, without an API key. It is currently a limited preview, and business sign-in is limited to selected partners.
• Apps SDK: Build ChatGPT apps with the Apps SDK and MCP. This guide did not check its content page by page, so see OpenAI’s Building with the Apps SDK.
• Pro and eligible Business plans have promotional API credits, separate from ChatGPT credits, and they need to be linked to an API organization (2026-10-28 release notes).

Source: Using Codex with your ChatGPT plan, Codex CLI, Codex IDE extension, Codex Cloud, Quickstart, Pricing, Models, Web search, Memories, DevDay 2026, API Quickstart, Using the latest model, ChatGPT release notes (queried 2026-11-06)

Privacy and Data Controls

I recommend everyone read this chapter. The key point: personal plans use your content to improve models by default, and you have to turn that off yourself, while business plans do not use it by default.

Training data options

• Personal plans (Free, Plus, Pro) do use content to improve models by default.
• There are two ways to turn it off, and both work for ChatGPT conversations and Codex tasks: turn off Settings → Data controls → Improve the model for everyone, or use “Do not train on my content” in the Privacy Portal.
• Turning off training does not delete or hide your chats.
• Business, Enterprise, Edu, Healthcare, and the API do not use content for training by default.

Exceptions that are easy to miss

1. If you give a thumbs up or thumbs down on a reply and submit feedback, the whole related conversation may be used for training, even if you have already opted out of training.
2. Codex has its own separate “Include environments” setting, and changing the ChatGPT setting does not carry over.
3. Inside a Trusted Access session (for example Daybreak), even with training turned off, interaction content may still be used to strengthen defenses against cyber abuse.
4. Voice has its own training sharing control.
5. Conversations with OpenAI Support may be used for improvement when Improve is on.

Temporary Chat

• It does not enter history, does not create or update memory, and is not used for training while it stays temporary. A copy may be kept for 30 days for safety purposes.
• From 2026-08-27 you can choose Personalized (can use memory, custom instructions, and plugins; decided before the chat starts and cannot be changed afterward) or Unpersonalized (the default).
• You can “Save” it afterward as a regular chat, which then follows your account settings.
• It shows no ads. Edits to writing blocks and code blocks inside Temporary Chat may not be kept.

Locked chats

Launched 2026-10-20: opened with a 4-digit PIN or device authentication, and hidden from history, pins, and search. It does not affect retention and training settings, and scheduled tasks do not support locked chats.

Exporting data

• Path: Settings → Data controls → Export data → Confirm export, or the Privacy Portal.
• Free, Go, Plus, Pro, and eligible Edu can use it. Business, Enterprise, and Healthcare do not offer self-service export.
• The notification is sent by email or SMS, in up to 7 days, and the download link expires after 24 hours.
• The ZIP contains chat history and other account data. Deleted chats cannot be recovered.
• If you are going to merge into a Business workspace, export before merging.

Deleting your account

Personal plans can delete from settings on their own, and deletion is permanent. Managed workspaces cannot self-delete. See Chapter 6 for the trap in deleting memory.

Privacy Center and Security history

• Privacy Center (2026-09-21): Brings together chat privacy, memory, personalization, data use, connected apps, and account security. It is at the web account menu → Help → Privacy Center.
• Security history (2026-09-25): Settings → Security and login lets you see recent sign-ins and security setting changes.

Search and location

Search sends the rewritten query to search partners, approximate location comes from your IP, and precise location is off by default. See Chapter 4 for details.

Ads

Free and Go may show ads. OpenAI says you can turn off ad personalization and clear ad data, and ads do not appear next to sensitive topics such as health, mental health, and politics. For ad details, rely on OpenAI’s Ads in ChatGPT page.

Suggested order of settings

1. First decide whether to let OpenAI train on your content, and turn it off first if you need to.
2. Check that Memory and Reference chat history are what you expect.
3. For sensitive topics, use Temporary Chat (Unpersonalized), or use Locked chats.
4. Before enabling Work’s cloud browser or Dots, read the prompt injection advice in Chapter 7.
5. If you use Codex, remember to check “Include environments” separately.

Source: Data controls in ChatGPT, How your data is used to improve model performance, Temporary chat in ChatGPT, Exporting your ChatGPT history and data, Scheduled tasks in ChatGPT, Ads in ChatGPT, the 2026-09-21, 2026-09-25, and 2026-10-20 entries in the ChatGPT release notes (queried 2026-11-06)

Prompting Tips and FAQ

OpenAI’s prompting advice is actually quite plain: state the result you want clearly, then add the necessary context and boundaries. This chapter organizes OpenAI’s prompting guides, along with the limits you run into most often in use.

OpenAI’s basic framework

A prompt can be a question, an instruction, or a goal. These four parts are the useful ones, and you include only the ones you need:

A few principles:

• Describe the result rather than the steps. Describe the process only when the process itself matters.
• Add context: attach files, attach screenshots and point out the area, use search when you want sources, share files with a project, and call a plugin with @. Preferences that span conversations go in custom instructions, and one-off details go in the prompt.
• Examples of boundaries: keep approved dates and budget figures; use only the sources provided, and flag gaps rather than guessing; draft the content but do not send it.
• For important work, ask it to check itself at the end (for example every action item has an owner and a deadline, and claims that cannot be verified are flagged), and then review it yourself once more.
• Use follow-ups instead of starting over: make the opening more direct, move recommendations to the front, add sources, change the approach.
• In Codex work you can Steer (add to the task currently running) or Queue (line it up for the next round). In the CLI, Enter is steer and Tab is queue.

General advice (from the prompt engineering help page): be clear and specific, iterate and refine, specify tone with adjectives, break complex tasks into smaller ones, and you can ask for several options.

Prompting tips for Work

• Include only relevant sources, and narrow the date range.
• Define the audience, format, and length.
• Separate required work from optional polish.
• When the method matters, ask for a plan first.
• Require approval before sending, publishing, or changing information that others rely on.
• When it does too much, stop it or narrow it down.
• For recurring tasks, tune the prompt in a normal chat before scheduling.

Prompting tips for Codex

• Point to the behavior, the relevant code or reproduction steps, the constraints, and how to verify.
• Investigate first with /plan.
• When fixing a bug: reproduce first, propose the smallest patch, add a regression test when feasible, and re-run the reproduction.
• Ask it to report which commands it ran and the results.

OpenAI’s prompting page has no GPT-6-specific techniques. Within the scope of this guide’s research, I also did not find a complete GPT-6 prompting guide, so I do not invent model-specific wording.

Frequently asked questions

A practical workflow

1. First try your phrasing in Chat with Instant or Medium to confirm the direction.
2. Turn on Search when you need sources, and use Deep research when you need a complete report.
3. Hand off to Work when you need a deliverable, and write the outcome, constraints, and review criteria clearly in the prompt.
4. Write settled preferences into custom instructions or project instructions, so you do not rewrite them each time.

Source: Prompting, Long-running work, Use ChatGPT, How do I create a good prompt for an AI model, Prompt engineering best practices for ChatGPT, GPT-6 and other models in ChatGPT, About ChatGPT Pro tiers, ChatGPT Free Tier FAQ, Using the latest model (queried 2026-11-06)

How ChatGPT Is Positioned Against Claude and Gemini

This chapter only covers product structure and plan differences that have an official page behind them. It does not compare “which model is smarter”. Each vendor’s benchmarks are published by that vendor itself, and this guide did not obtain reliable, directly comparable official data, so test model quality on your own work.

Plan and price comparison

The table below is what each vendor’s official pricing page showed on 2026-11-06. NT$ figures are reference prices, and USD figures come from each vendor’s pages:

Gemini’s privacy defaults, ads, and fuller model specifications are not included in the comparison, because this guide did not obtain official pages to support them.

A few observations with sources behind them

1. ChatGPT puts chat, Work deliverables, Codex, voice, images, and scheduling in one account. OpenAI calls these the three modes Chat, Work, and Codex, and announced that from 2026-12 the Chat and Codex quotas will be merged. This is a fact about product structure, not a quality ranking.
2. The paid ladder is the finest-grained: ChatGPT is Go $8, Plus $20, Pro $100 / $200 / $500; Claude is $20, $100, $200; Google’s Taiwan page has NT$165, 650, 3,300, and 6,500 (reference prices). The price bands are very close, and the difference is in how quota is counted and which features are included.
3. Usage rules differ by vendor, and all of them are changing: ChatGPT is “Instant unlimited, reasoning and tools each limited, Work and Codex measured in tokens or credits”; Claude counts in multiples per 5 hours and per week; Google uses compute units, resetting every 5 hours up to a weekly cap. When comparing, look at whether your actual workflow will hit the limit, and do not compare only the monthly fee.
4. The developer side maps directly: ChatGPT’s Codex (CLI, IDE, Cloud, desktop) and Claude Code are the same kind of product, both with a CLI, IDE extension, cloud tasks, and a project instruction file (Codex uses AGENTS.md, Claude Code uses CLAUDE.md). For Claude Code details, see this site’s Claude Code complete guide.
5. Privacy defaults: ChatGPT and Claude business plans both default to not training on your data. On personal plans, ChatGPT uses data for improvement by default and you can opt out (see Chapter 11), and Claude’s page says consumer plans can opt out.
6. Ads are a variable unique to ChatGPT: Free and Go have ads, and Plus and up do not. For the other two vendors, this guide did not find a corresponding official statement, so I draw no conclusion.

How to choose

• If your main work is conversation, search, and making slide and document deliverables, and you also want voice and images: start with Plus, and look at Pro if the quota feels insufficient.
• If you mainly write code: both Codex and Claude Code are worth a real trial round on your own project. The thing to compare is whether you hit the quota wall and whether the workflow feels smooth, not leaderboards.
• For team use: Business does not train on your data by default, and Standard and Premium seats can be mixed.

Whichever vendor you pick, quotas and plans can change within a month or two. Before you commit, go back to the official pages and confirm the current terms.

Source: ChatGPT pricing, About ChatGPT Pro tiers, Understanding ChatGPT and Codex usage limits, Claude pricing, Google subscription plans (as displayed in Taiwan) (queried 2026-11-06)