锣鼓喧天、AI hype 拉满的这一年，终于等来 Anthropic 的招股书。招股书里最核心的增长引擎是 Claude Code：头部企业客户每月支出是普通用户的 20 倍，9 月 Anthropic 代码库 90% 以上的代码由它写出。
但同一份文件承认，5% 到 7% 的收入来自“违反使用规定”的客户。按前三季度约 331 亿美元营收折算，就是 17 亿到 23 亿美元。用着 Claude Code、续完年费、然后被封号一分不退的中国开发者，一看就知道这里头有谁的钱。
Claude Code：招股书里的核心产品，开发者用钱包投出来的
Anthropic 周一提交 S-1 文件，计划以股票代码 ANT 登陆纳斯达克。文件将 Claude Code 描绘为推动公司企业业务增长的核心产品，并把它放在了整个企业增长故事的正中央。
S-1 文件里有一段对企业客户的拆解：第三季度，没有使用 Claude Code 的 Claude Enterprise 用户平均每月支出不到 75 美元；使用 Claude Code 的企业用户中，消费最高的 5% 平均支出约为这一数字的 20 倍。Claude Code 几乎就是一台客单价拉升引擎。
Claude Code 已经从提供代码建议，发展到直接编写和运行代码，并能在较长时间内自主处理复杂的多步骤工程任务。工程师因此越来越多地负责指挥和审查 Claude 的工作，而不必逐行编写代码。截至 2026 年 9 月 30 日的三个月里，每位工程师平均每天为公司代码库贡献的代码量，超过截至 2024 年 12 月 31 日年度日均代码量的 45 倍。2026 年 9 月，公司代码库中超过 90% 的贡献由 Claude 编写。
Anthropic 的这个文件还展现了一套典型的“先落地、再扩张”策略，而 Claude Code 承担了打开局面的角色：一名工程师开始使用 Claude Code，觉得好用，便在公司内部推广。之后 Anthropic 的销售团队才会介入。文件称，到那时，“Claude 往往已经在使用中了”。这意味着 Claude Code 不是销售卖出去的产品，而是自己长进企业里的产品——获客成本由工程师的口碑承担，销售只需收割。
Anthropic 自己也不掩饰对它的偏爱，称 Claude Code 是“虚拟协作者最鲜明的代表，也是我们认为最具体地展现了自主 AI 工作形态的产品”、“发布后的几个月内，自主编程就成为衡量前沿 AI 进展的重要参照”。
虽然“软件开发者是 Claude 的重要推广者“，但这种扩散也越过了工程团队。
招股书称，非技术团队越来越多地使用 Claude Code 和 Claude Cowork 管理自己的工作流。Anthropic 对 Cowork 的描述延续了它对 Claude Code 的定位：“所有数字化工作最终都可以归结为代码。”
Cowork 将 Claude Code 的智能体架构延伸给知识工作者：用户描述任务后，它可以读取文件、操作应用，在桌面上执行多步骤工作流，同时把底层代码隐藏起来。Anthropic 表示他们
“还会针对特定职业和行业开发专用应用，在 Claude Code 的基础上继续拓展，加快 Claude 进入新的知识工作领域，并扩大其在企业中的应用范围。”
这也意味着 Claude Code 从编程工具变成了 Anthropic 进军更广泛工作场景的技术起点。截至 2026 年 9 月 30 日，Cowork 推出研究预览后的同期采用速度，已经超过 Claude Code 推出研究预览后的同期表现。
Anthropic 还在把支撑 Claude Code 的 Harness 变成开发者可以购买和使用的平台能力。文件用一句话概括这一思路：“单靠一个模型，还构不成 Agent。”
开发者可以通过 Agent SDK 在自己的服务器上运行 Anthropic 的 Harness，也可以付费让 Anthropic 托管运行。这让 Anthropic 越来越接近那些基于其模型构建编程智能体的公司，而 S-1 在竞争格局里只提了 Google 和 OpenAI，没提 Cursor、Replit 和 Cognition。
Claude Code 拉进来的人，靠新模型持续掏钱
Claude Code 把工程师带进公司，模型更新则给了企业继续加码的理由。Anthropic 在 S-1 里写，企业客户会在新模型发布后尽快转向最新、能力最强的版本，把它用在价值最高的工作上。文件还写道：“每次发布新版本后，我们都看到大多数客户会使用模型的最新版本。”换句话说，Claude Code 打开的入口，最终通向的是 Anthropic 整个模型家族。
这套逻辑已经直接体现在收入里。Anthropic 称，截至 2026 年 9 月 30 日，过去六个月内发布的模型贡献了 Anthropic 第三季度 80% 以上的收入。2026 年 4 月至 9 月，Anthropic 发布了 8 款面向广泛用户开放的 Claude 模型，以及 3 款仅限特定用户访问的 Mythos 模型。
招股书还披露，2026 年前九个月，Anthropic 合计推出 250 项新模型版本、平台功能和应用（这是三类发布的合计数，文件没有分别列出各类数量）。
公司称，客户的支出与新模型的发布频率及模型能力密切相关。每一次性能提升，都会让客户有更多理由使用 Claude；能力更强的模型也能吸引新客户，并促使现有客户增加使用量、提高支出。
这些收入又能为下一轮研发提供资金。Anthropic 在文件中表示，公司依靠不断增长的收入，支付研发、算力和技术人才等成本，以构建能力更强的模型。模型能力提升后，可能进一步吸引新客户并带来更多支出，由此形成一个循环：每一代 Claude 都在为下一代提供资金。
还有一个让华尔街兴奋的数字：900%。
这是 Anthropic 截至第三季度的净收入留存率。一年前就在付钱的客户，今年平均多掏了九倍。年付超过 10 万美元的客户中，超过一半同时使用了 Claude 平台和 Claude Code 或 Cowork 这类第一方应用，2025 年底这个比例只有约 30%。
企业客户的数量也在同步增长。截至 9 月 30 日，年付费超过 10 万美元的客户超过 11,900 家，一年前还不到 950 家；年付超过 100 万美元的超过 1,900 家；年付超过 1000 万美元的约有 250 家。Anthropic 的总企业客户数超过 30 万。
收费方式让这种依赖更明显。Anthropic 2026 年上半年的按量计费收入达到 131 亿美元，占总收入约 80%；2024 年这一数字还是 2.88 亿美元。客户用得越多，账单越大；模型越能接手高价值任务，Anthropic 就越有机会把一次 Claude Code 试用，变成企业里持续增长的模型调用量。
那 5% 到 7%：封号不退费的钱，写进了营收
招股书里还有另一段披露，被放到了收入里。
截至 2026 年 9 月 30 日的九个月内，约 5% 到 7% 的收入来自违反 Anthropic 使用政策或服务条款的客户，公司预计随着检测和执法力度加强，这一比例会下降。这部分收入来自“被发现后失去访问权限的客户”。
截图来源：S-1 第 145 页
说得更直白一点就是每收 1 美元，就有 5 到 7 美分来自公司不想留在平台上、打算清理掉的用户。按前三季度约 331 亿美元营收折算，这笔钱在 17 亿到 23 亿美元之间。
关键细节在于收入确认的顺序：Anthropic 承认，收入是在服务交付时确认的，然后才限制或终止账户。先确认收入，再执行风控——这个顺序，比任何转述都更接近被封号用户真正关心的那个机制。不是“违规所以没有收入”，而是“收入已经记上了，然后才封的号”。
招股书没有拆分这笔钱的来源：多少是 API 滥用，多少是蒸馏行为，多少是跨区订阅后被封禁的个人开发者，多少是刚续完年费就被风控封掉的团队账号。但 Claude Code 的中国开发者在 GitHub 上 anthropics/claude-code 仓库的 issue 记录比任何转述都原始：
- Issue #37057：用户 3 月 16 日订阅 Claude Pro，17 日升级 Max 5x，18 日升级 Max 20x 并支付约 209 美元，同一天账号被禁用，无法访问 Claude.ai 和 Claude Code，申诉无回复，要求按比例退款。
- Issue #91411：用户 9 月 1 日支付 249 美元购买 Max 20x，用 Claude Code 做正常的软件开发工作，第二天整个账号被禁用，没有具体原因、没有警告、没有自动退款。用户原话是：“收了 249 美元，第二天封号，没有具体理由，没有立即退款，和欺骗性计费没有区别。”
- Issue #41252：用户用 Max 5x 礼品订阅，在正常使用 claude.ai 和 Claude Code 的情况下账号被禁用，申诉一周无回复，100 美元无法使用也无退款路径。
这些人用的正是 Claude Code。他们续的年费、充的额度、跑掉的 token，在封号那一刻没有退回，而是留在了 Anthropic 的营收里，如今写进了给华尔街看的文件，被归入“违反政策客户”贡献的收入。
Claude Code 是 Anthropic 递给华尔街的招牌。那 5% 到 7%，很多是那些 Claude Code 年费交了、号没了、钱也没了的中国开发者。
这笔钱也成了 Dario 口袋里的一部分。
S-1 首次披露了 Anthropic 的股权结构：七位创始人持股数量大致相当，锁定期到期后，纸面财富会变成真实的财富。但控制权不会旁落——公司发行的“Class F”创始人股份，让七个人合计拥有相当于流通股 50.1% 的投票权。
有网友的评论很中肯：“Anthropic 可能会是全球历史上最大 IPO，但同时也可能是全球最不要脸的公司”。
参考链接：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It's the year of the raging, al-hype, waiting for the Anthropic book. The most central growth engine in the book is Claude Code: the head business customer spends 20 times the average user per month, and in September, the Anthropic Code Collection, more than 90% of the code is written by it.
But the same document recognizes that between 5 and 7 percent of the income comes from a client who violates the rules. Based on approximately $33.1 billion received in the previous three quarters, it was $1.7 billion to $2.3 billion. With Claude Code, a Chinese developer who finishes his annual fee and is then sealed off, he knows who the money is.
Claude Code: The core product in the book, the developer dropped it in his wallet.
Anthropic submits the S-1 file on Monday and plans to land in NASDAQ with the stock code ANT. The document depicts Claude Code as the core product that drives corporate business growth and puts it right in the centre of the whole business growth story.
The S-1 document contains a section on the dismantling of business clients: in the third quarter, Claude Enterprise users who did not use Claude Code spent less than $75 per month on average, and the highest 5 per cent of business users who used Claude Code spent about 20 times the average. Claude Code is almost a single-price lift engine.
Claude Code has evolved from providing code advice to directly preparing and operating codes and is able to handle complex multi-step engineering tasks on its own over a longer period of time. Engineers are therefore increasingly responsible for directing and reviewing Claude without having to write code by line. In the three months of September 30, 2026, each engineer contributed an average of 45 times the average number of codes per day to the company's code repository as of December 31, 2024. In September 2026, more than 90% of the contributions in the company code library were written by Claude.
This document by Anthropic also presents a typical “first-down, then-extension” strategy, and Claude Code has taken on the role of opening the door: an engineer started using Claude Code, which was useful and then promoted within the company. After the Anthropic sales team intervened. The document states that by that time, “Claude is often already in use”. Which means Claude Code wasn't selling the product that he sold, but the product that he grew into. - The cost of the take-over is borne by the reputation of the engineer, and the sale requires only harvesting.
Anthropic himself does not hide his preference for it, saying that Claude Code is “the most visible representative of virtual collaborators and, in our view, the product of the most concrete demonstration of an autonomous AI working pattern”, and that in the months following publish, autonomous programming becomes an important reference for measuring progress at the front line of AI.
Although “software developers are important promoters of Claude”, this proliferation also transcends engineering teams.
According to the book, non-technical teams are increasingly using Claude Code and Claude Work to manage their own workflow. Anthropic describes Cowork as a continuation of its positioning for Claude Code: "All digitization can end up as code."
Cowork extends the AI agent structure of Claude Code to knowledge workers: after the user describes the task, it can read files, operate applications, execute multiple steps workflow on the desktop, and hide the bottom code. Anthropic means them.
“Special applications will also be developed for specific occupations and industries, building on Claude Code to accelerate Claude's entry into new areas of knowledge and expand its application in enterprises.”
It also means that Claude Code has moved from a programming tool to a technical starting point for anthropic to a broader work scene. As at 30 September 2026, Cowork had introduced a research preview at a rate that exceeded Claude Code at the same time.
Anthropic is also transforming the Harness that supports Claude Code into a platform that developers can buy and use. The paper summed up the line in a single sentence: “Agent cannot be created by a single model”.
Developers can run Anthony's Harness on their own server via Agent SDK, or paid for Anthropic hosting. This brought Anthropic closer to the companies that built the AI agent program based on its model, while S-1 mentioned Google and Open AI in the competition pattern, without reference to Cursor, Replit and Cognition.
Claude Code, the guy who pulled in, he just kept taking it from the new model. Money.
Claude Code brought the engineers into the company, and the model upgrade gave the company reason to continue adding. Anthropic wrote in S-1 that business clients will move to the latest and most capable version of the new model publish as soon as possible, using it for the most valuable work. The document also states: “Every time we see a new version of publish, we see that most customers will use the latest version of the model.” In other words, Claude Code opened the entrance to the whole model family of Anthropic.
This logic is already reflected directly in income. According to Anthropic, as of 30 September 2026, the model of publish over the past six months contributed 80% of Anthropic ' s income in the third quarter. From April to September 2026, Anthropic publish made eight Claude models open to a wide range of users and three Mythos models limited to specific users.
The equity book also disclosed that in the first nine months of 2026, Anthropic had launched a total of 250 new model versions, platform functions and applications (this is the sum of three types of publish, which were not separately listed in the document).
According to the company, the client ' s expenditure is closely related to the frequency and modelling capability of the new model publish. Each one-time upgrade would give clients more reason to use Claude; a more capable model would also attract new clients and encourage existing clients to increase usage and expenditure.
These revenues can in turn finance the next R&D round. In the paper, Anthropic states that companies rely on growing revenue to cover the costs of R & D, numeracy and skilled personnel to build more capable models. The enhanced capacity of the model may further attract new clients and lead to more expenditure, resulting in a cycle in which each generation Claude is funding the next generation.
There's another number that turns Wall Street on: 900%.
This is the net income retention rate of Anthropic up to the third quarter. Clients who paid a year ago paid nine times more this year on average. More than half of the clients paying more than $100,000 per year used both Claude Platform and first-party applications such as Claude Code or Cowork, which was only about 30 per cent at the end of 2025.
The number of business clients is also growing simultaneously. As at 30 September, paid clients over $100,000 a year exceeded 11,900 and less than 950 a year ago; more than 1,900 had paid more than $1 million a year; and about 250 had paid more than $10 million a year. Anthropic has more than 300,000 corporate clients.
This dependence is made more evident by the fee modality. In the first half of 2026, volume revenue amounted to $13.1 billion, or about 80 per cent of total income; in 2024, the figure was $288 million. The more clients use it, the bigger the bill; the more the model takes over high-value tasks, the more Anthony will have the opportunity to turn a Claude Code trial into a model that continues to grow.
So, 5% to 7%: No refund for the seal, written in the collection.
There was another disclosure in the equity book, which was placed in the income.
In the nine months of September 30, 2026, about 5 to 7 per cent of the revenues were generated by customers who violated the policy or service terms of the Anthropic use, a proportion that companies expected to reduce with increased detection and enforcement. This revenue is derived from “a customer who lost access to permission upon discovery”.
Source: S-1 145
Let's be clear: for every dollar, five to seven cents comes from a user that the company doesn't want to stay on the platform and try to clean up. This amount is between $1.7 billion and $2.3 billion.
The key detail is the order in which income is recognized: Anthropic acknowledges that income is recognized at the time of service delivery before the account is restricted or terminated. First, revenue is recognized and then wind control is carried out — a sequence that is closer than any recitation to the mechanism of genuine interest to the users of the sealed name. It is not “breaches so that there is no income”, but “income is recorded and then sealed”.
split is not the source of this money in the book: the API abuse, the distillation, the individual developers who were banned after cross-zonal subscriptions, and the team account that was blocked after the annual fee had been renewed. However, a Chinese developer from Claude Code on GitHub has an anthropics/claude-code warehouse with an assue record that is more original than anything else:
- Issue #37057: User subscribes to Claude Pro on March 16, upgrades Max 5x17 to Max 5x, upgrades Max 20x on March 18, and pays approximately US$ 209, the same day account number is disabled and unable to access Claude.ai and Claude Code, without a response and requests a proportional refund.
- Issue #91411: The user paid $249 on 1 September to purchase Max 20x, do normal software development with Claude Code, and the entire account was disabled the next day for no specific reason, no warning, no automatic refund. According to the user: “A charge of US$ 249, with the next day's seal, without a specific reason, without an immediate refund and without distinction from fraudulent charges.”
- Issue #41252: Users subscribe with Max 5x gifts, account numbers are disabled with normal use of claude.ai and Claude Code, complaints are not answered in a week, $100 cannot be used or refunded.
These people are using Claude Code. Their annual fee, the amount of quota, the runaway token, did not return the seal at the moment, but left it in the collection of Anthropic, and is now included in a document for Wall Street, which is classified as income from a “breacher”.
Claude Code is a sign from Anthropic to Wall Street. That's between 5 and 7 percent, and many of those Chinese developers that Claude Code has paid an annual fee, no., no money.
The money became part of Dario's pocket.
S-1 revealed for the first time the stock structure of Anthropic: the seven founders held approximately the same number of shares, and after the lock expires periodically, paper wealth becomes real wealth. But control does not go side-by-side - the company's “Class F” founders' shares issued, giving seven persons a combined vote equivalent to 50.1 per cent of the circulation stock.
The comment made by the netizens is very pertinent: “Anthropic may be the largest IPO in global history, but it may also be the world's most disgraceful company.”
References:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.