引子 · 你的订阅不止那个聊天框Intro · Your Subscription Is More Than That Chat Box.
你为 ChatGPT 付了钱，几乎所有对话都在网页那个输入框里。这本书讲另外那部分 —— 同一份订阅还打开了模型选择器、Memory 与 Projects、Plugins / Apps、自建 GPT、Deep Research、交付长任务的 Work、能上网动手的 agent、写码的 Codex，以及图像、语音与 Sora 边界。不是产品清单，是一把筛子：遇到哪类活，该开哪一个。You pay for ChatGPT, and nearly every conversation happens in that one web input box. This book is about the rest — the same subscription also opens the model picker, Memory and Projects, Plugins / Apps, custom GPTs, Deep Research, Work for longer deliverables, an agent that acts on the web, Codex for code, plus images, voice, and the Sora boundary. Not a feature list; a filter — which one to open for which kind of work.
你大概是这样用 ChatGPT 的：打开网页或 App，在那个输入框里打字，读回答，再打字。一年下来几千条对话，几乎全在同一个框、同一个默认模型上。Here is probably how you use ChatGPT: open the web page or the app, type into that one input box, read the reply, type again. Thousands of conversations over a year, nearly all in the same box and on the same default model.那个框很好用，好用到你忘了「ChatGPT」只是 OpenAI 给你的一个产品。同一份订阅后面还排着一串别的：藏在选择器里的几个模型、自动记事的 Memory、把文件钉住的 Projects、把 workflow 与数据连接装进来的 Plugins / Apps、能自建的 GPT、出带引用报告的 Deep Research、交付长任务成品的 Work、有自己浏览器的 agent、跑在本机和编辑器里的 Codex，还有图像、语音，以及已经变动的视频 / Sora 边界。它们不长得像同一个东西 —— 有的是网页、有的是手机 App、有的干脆是命令行。这本书讲的，就是这些你付了钱、却没点开过的产品。That box is good — good enough that you have forgotten ChatGPT is just one of the products OpenAI hands you. Behind the same subscription sits a row of others: models hidden in the picker, Memory that takes notes, Projects that pin files, Plugins / Apps that package workflows and connect data, custom GPTs, Deep Research for cited reports, Work for longer deliverables, an agent with its own browser, Codex on your machine and in the editor, plus images, voice, and the shifting video / Sora boundary. They do not even look like the same thing — some are web pages, some are phone apps, one is a command line. This book is about the products you have paid for and never opened.
— I
你买的是 ChatGPT，到手的是一整套You Bought ChatGPT, You Got a Whole Set.
你订的是一个聊天框，能用的是一柜子产品。You subscribed to a chat box and unlocked a cabinet of products.截至 2026-07-10，ChatGPT 的定价页已经把 GPT-5.5 Instant、GPT-5.6、Work、Plugins、Memory、Projects、Codex、Deep Research、图像和语音放在同一张计划表里比较。1注 1Note 1ChatGPT 套餐与价格页 —— 截至 2026-07-10，页面列出 Free、Go、Plus、Pro、Business、Enterprise，并比较 GPT-5.5 Instant、GPT-5.6 Sol / Sol Pro、Work、Plugins、Codex、Deep Research、Projects、图像和语音等能力；具体价格、额度和地区可用性以页面与账号内显示为准。ChatGPT Plans & Pricing — as of 2026-07-10, the page lists Free, Go, Plus, Pro, Business, and Enterprise, comparing GPT-5.5 Instant, GPT-5.6 Sol / Sol Pro, Work, Plugins, Codex, Deep Research, Projects, images, voice, and more; exact price, allowance, and regional availability should be checked on the page and in-account UI. 具体价格、额度和地区可用性要看页面和你账号内实际显示，但这件事已经足够说明：你买的不是单个聊天框，而是一组按计划开放的入口。标准 Chat 里，Instant 仍由 GPT-5.5 驱动，Medium / High 等推理入口与 Pro 才转到 GPT-5.6 Sol / Sol Pro；2注 2Note 2OpenAI 帮助中心 · GPT-5.6 in ChatGPT —— 截至 2026-07-10，GPT-5.6 正向符合资格的计划逐步推出；GPT-5.5 Instant 仍是默认快答，GPT-5.6 Sol 驱动 Medium / High / Extra High，Sol Pro 驱动 Pro；Terra / Luna 不可在标准 Chat 对话中选择。OpenAI Help · GPT-5.6 in ChatGPT — as of 2026-07-10, GPT-5.6 is gradually rolling out to eligible plans; GPT-5.5 Instant remains the default for fast responses, GPT-5.6 Sol powers Medium / High / Extra High, Sol Pro powers Pro, and Terra / Luna are not selectable in standard Chat conversations. Projects 和 Memory 负责上下文，Plugins / Apps 负责 workflow 与外部资料，Work 负责长任务成品，agent 负责具体网页行动，Codex 负责软件开发。As of 2026-07-10, ChatGPT's pricing page compares GPT-5.5 Instant, GPT-5.6, Work, Plugins, Memory, Projects, Codex, Deep Research, images, and voice on the same plan table.1注 1Note 1ChatGPT 套餐与价格页 —— 截至 2026-07-10，页面列出 Free、Go、Plus、Pro、Business、Enterprise，并比较 GPT-5.5 Instant、GPT-5.6 Sol / Sol Pro、Work、Plugins、Codex、Deep Research、Projects、图像和语音等能力；具体价格、额度和地区可用性以页面与账号内显示为准。ChatGPT Plans & Pricing — as of 2026-07-10, the page lists Free, Go, Plus, Pro, Business, and Enterprise, comparing GPT-5.5 Instant, GPT-5.6 Sol / Sol Pro, Work, Plugins, Codex, Deep Research, Projects, images, voice, and more; exact price, allowance, and regional availability should be checked on the page and in-account UI. Exact price, allowance, and regional availability belong to the page and your account UI, but that alone proves the point: you did not buy one chat box; you bought a set of entrances opened by plan. In standard Chat, Instant still runs on GPT-5.5, while Medium / High reasoning and Pro move to GPT-5.6 Sol / Sol Pro;2注 2Note 2OpenAI 帮助中心 · GPT-5.6 in ChatGPT —— 截至 2026-07-10，GPT-5.6 正向符合资格的计划逐步推出；GPT-5.5 Instant 仍是默认快答，GPT-5.6 Sol 驱动 Medium / High / Extra High，Sol Pro 驱动 Pro；Terra / Luna 不可在标准 Chat 对话中选择。OpenAI Help · GPT-5.6 in ChatGPT — as of 2026-07-10, GPT-5.6 is gradually rolling out to eligible plans; GPT-5.5 Instant remains the default for fast responses, GPT-5.6 Sol powers Medium / High / Extra High, Sol Pro powers Pro, and Terra / Luna are not selectable in standard Chat conversations. Projects and Memory hold context, Plugins / Apps package workflows and connect external material, Work produces longer deliverables, the agent handles concrete web actions, and Codex handles software development.问题不是你懒。问题是产品页只告诉你「能做什么」，从不告诉你「在你这种活里，这部分值不值得花十分钟」。OpenAI 的产品还格外散 —— Codex 要在命令行里装，平台又是另一个网站，每个面长得都不一样。每多一个面，就多一次「要不要学」的决定，而你手头有正事，于是默认全部跳过，退回那个最熟的框。The problem is not that you are lazy. The problem is the product page only tells you what it can do, never whether, given your kind of work, this part is worth ten minutes. OpenAI's products are unusually scattered, too — Codex installs in a command line, the platform is yet another website, and no two surfaces look alike. Each extra surface is one more should-I-learn-this decision, and you have real work to do, so you skip them all and retreat to the box you know best.所以广度不是会员的福利，是会员的税。这本书想做的，是替你把这笔税里真正划算的那几样挑出来。So breadth is not the membership's perk — it is its tax. What this book tries to do is pick out the few items in that tax that actually pay off.
— 01 / 04 · MAP · 总览
OpenAI 产品入口OpenAI product entrances
同一份订阅不是一个框，而是一组入口。The subscription is not one box, but a set of entrances.
聊天框→上下文→查证 / 动手→代码 / 表达chat→context→research / action→code / media
先把背景、文件和外部资料放对位置。Put background, files, and outside material in the right place first.
偏好→文件→外部资料preferences→files→external material
— 03 / 04 · ACTION · 执行
Deep Research · AgentDeep Research · Agent
一个负责查清楚，一个负责在网页里动手。One finds out; the other acts inside a browser.
问题→引用报告→浏览器动作question→cited report→browser action
— 04 / 04 · MAKE · 产出
Codex · 图像 · 语音 · Sora 边界Codex · Images · Voice · the Sora boundary
代码、图像与语音仍可交付；Sora 是已退役边界。Code, images, and voice remain deliverable; Sora is a retired boundary.
代码→图像→双向语音 · 视频入口已停用code→image→duplex voice · video entrance retired
01 / 04
— MAP · 单图
一件事，一个入口One task, one entrance
把常见任务先分流，再决定该打开哪个 OpenAI 产品。Route the task first, then choose the OpenAI surface.
任务→入口→交付task→entrance→output
— II
这是一把筛子，不是清单This Is a Filter, Not a Catalog.
我替你把长尾砍掉，只留能覆盖八成场景的那两成。I cut the long tail for you and keep the 20% that covers 80% of the cases.每一项核心能力，这本书只回答三个问题：省什么、什么场景用、什么时候别用。不教你从零点哪个按钮 —— OpenAI 的界面三个月一变，按钮会改版，判断不会。给你的是一把尺子，量你自己的活，不是一条让你照走的路。For each core capability, this book answers only three questions: what it saves, when to use it, and when not to. It will not teach you which button to click from scratch — OpenAI's interface changes every few months, buttons get redesigned, judgment does not. What you get is a ruler to measure your own work against, not a path to follow.留下的两成，就是后面的章节：挑模型、用 Memory 与 Projects 记住、Plugins / Apps、自建 GPT、Deep Research、图像、语音与 Sora 边界、Work / agent、Codex，最后是怎么选订阅档。The 20% that stays is the chapters ahead: picking a model, remembering with Memory and Projects, Plugins / Apps, building a GPT, Deep Research, images, voice, and the Sora boundary, Work / agent, Codex — and finally how to choose a tier.砍掉的是什么，也说清楚：API 平台、Agents SDK、批量与微调、在 Azure 上的部署、企业的 SCIM 与合规后台 —— 那些是给开发者和 IT 的，不是给一个被产品广度淹没的付费用户的。尤其那条最容易踩的线：订阅不含 API 额度，platform.openai.com 是另一套账户、另一条账单。3注 3Note 3OpenAI 帮助中心 · ChatGPT 与 Platform 的计费是分开的 —— 订阅（Plus / Pro / Business 等）不附带 API 额度，API 在 platform.openai.com 单独按 token 计费。OpenAI Help · ChatGPT and Platform are billed separately — a subscription (Plus / Pro / Business, etc.) does not include API credit; the API is billed per token on platform.openai.com.它们很重要，只是不在你这十分钟里。What gets cut is named too: the API platform, the Agents SDK, batch and fine-tuning, deploying on Azure, enterprise SCIM and the compliance console — those are for developers and IT, not for a paying user drowning in product breadth. Especially the easiest trap: a subscription includes no API credit; platform.openai.com is a separate account and a separate bill.3注 3Note 3OpenAI 帮助中心 · ChatGPT 与 Platform 的计费是分开的 —— 订阅（Plus / Pro / Business 等）不附带 API 额度，API 在 platform.openai.com 单独按 token 计费。OpenAI Help · ChatGPT and Platform are billed separately — a subscription (Plus / Pro / Business, etc.) does not include API credit; the API is billed per token on platform.openai.com. They matter; they are just not in your ten minutes.
— III
选对入口，比追更高的档位重要The Right Entrance Beats a Higher Tier.
同一份订阅，能力早就在那里了。差的不是档位，是你有没有走对那扇门。Within one subscription, the capability is already there. What is missing is not a tier — it is whether you walked through the right door.记住这一句就够：选择器挑脑子、Projects 攒上下文、Deep Research 替你查、agent 替你做、Codex 替你写。同一段需求，丢进默认聊天是一问一答地试，丢进 Deep Research 是等十分钟换一份带引用的报告，交给 agent 是它在自己的浏览器里点完再回来交差。多数人卡住，不是因为档位太低，是因为所有活都塞进了同一个框。One line is enough: the picker picks the brain, Projects holds the context, Deep Research finds out, the agent does it, Codex writes it. The same request, dropped into the default chat, is trial-and-error; dropped into Deep Research, you wait ten minutes for a cited report; handed to the agent, it clicks through its own browser and comes back done. Most people are stuck not because their tier is too low, but because every task goes into the same box.所以这本书不必从头读。不碰代码，就跳过「写码 · 从 Canvas 到 Codex」那一章 —— 把它单拎出来，就是为了让你能干净地跳过。要是只想先翻两章，翻「挑模型」和「记住」：前者是你每天都在用、却从没调过的开关，后者最快止住你每次重贴背景的损耗。So you do not have to read this front to back. If you never touch code, skip the chapter Write Code · From Canvas to Codex — it is a separate one precisely so you can skip it cleanly. If you only want two chapters first, read Pick a Model and Remember: the first is the switch you use every day and have never adjusted; the second is the fastest way to stop the bleed of re-pasting your background.动手之前先照一次镜子 —— 看看你现在到底把活都塞进了哪里：Before doing anything else, hold up a mirror — see where you have actually been cramming all your work:
01
翻记录，挑 5 件Scroll back, pick 5
打开你的 ChatGPT 历史，往回翻两周，挑出 5 件你让它做过、且不止做过一次的事。Open your ChatGPT history, scroll back two weeks, and pick 5 things you have asked it to do — more than once.
02
给每件标两列Tag each with two columns
列一张五行的表：事项 / 你当时用的产品 / 现在你觉得更该用的产品（默认 Chat、选择器换推理档、Projects、Plugins / Apps、GPT、Deep Research、Work、agent、Codex —— 先凭直觉填，读完再回来改）。Make a five-row table: task / the product you used then / the product you now think it belonged to (default Chat, a different reasoning level, Projects, Plugins / Apps, a GPT, Deep Research, Work, the agent, Codex — fill it on instinct now, come back and fix it after reading).
03
数一数有几行写着「默认聊天」Count how many rows say default chat
第二列大概率五行都是默认聊天、默认模型。这张表就是你读这本书的起点，也是读完之后的对照组。The second column is probably default chat on the default model for all five rows. That table is your starting point for this book — and your control group once you have finished it.
做完你会发现，不是 ChatGPT 少了什么能力，是你一直只站在一个入口。When you are done, you will see it is not that ChatGPT lacks some capability — it is that you have been standing at one entrance the whole time.
你订的是一个聊天框,
到手的是一整套产品.
You subscribed to one chat box,
you got a whole set of products.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return