文 | 奇点湃
10月24日，45分钟之内，Claude Code 最知名的两位成员在 X 上先后官宣离职。
一位是 Thariq Shihipar，Claude Code 团队的核心工程师，被开发者社区称作「解释之王」，skills 系统和上下文工程都和他有关。另一位是 Anthony Morris，Claude Code 桌面版的主创，他说自己在 Anthropic 的最后一天是10月23日。
两人都没说下一站去哪，告别信写得体面。Shihipar 还给前东家留了一句好话，说同事是他遇到过最有才华的一群人。
这一天，距离 Anthropic 计划中的 IPO 只有三周左右。这家公司的目标估值最高喊到过 2 万亿美元，有望刷新全球 IPO 纪录。路演还没开始，产品线上最出名的两张脸先走了。
01 两个金主，同时踩刹车
离职只是台面这条线。更早的10月5日，The Information 报道了另一件事：Anthropic 最大的两个企业客户，微软和 Meta，都在压缩内部员工对 Claude 的使用。
微软高管今年早些时候预估，公司内部使用 Anthropic 技术的年支出至少会到 10 亿美元。现在这个预算被砍掉了超过三分之一，员工被要求少用 Claude，改用 GitHub Copilot 和自研模型。
刀法更细的是额度。微软云与 AI 事业部有 6 万多名员工，此前每人每月的 AI 模型预算上限是 10 万美元，近期多数场景被压到约 1 万美元，一刀砍掉九成。有工程师因此抱怨，模型不再想用就用。
Meta 那边，内部使用 Claude Code 的员工从今年早些时候的约 6 万人降到约 3 万人。腰斩里有一部分是春季裁员的余波，Meta 当时裁掉了约 10% 的岗位。但更主要的原因是 Meta 在推自研工具：内部专用的 Meta Code 用户已超 3 万，对外测试的 Muse Code 也有 6000 多名员工在用。
有一个数字值得掰开看。用户腰斩之后，Meta 最近 28 天在 Claude Code 上仍花掉超过 1.05 亿美元。人走了一半，账单没怎么瘦。
微软砍的是内部办公和研发的支出，微软为外部客户采购 Claude 用于 Copilot 功能的那部分，原本预计每年至少 20 亿美元，据报仍在稳步上涨。同样，Meta 和微软依然通过各自的云平台向外部客户卖 Claude。
所以准确的表述是：金主没有退订 Claude，金主只是不再让自己的员工敞开了用。
02 贵，成了一种罪
为什么收缩？复盘微软的时间线，答案简单得有点残酷：Claude Code 太好用了。
2025年12月，微软给体验与设备部门的数千名工程师开通了 Claude Code。没人推广，没人动员，它靠口碑自己火了，头几个月 token 消耗量直接翻倍。
问题随之变成一道算术题。每个工程师每月的 token 成本在 500 到 2000 美元之间，一个团队加起来就是数百万美元。AI 工具第一次像云资源一样，需要被设预算、被审计、被限额。
这背后是 Claude 的价格位置。以2026年10月的官方定价看，Opus 主力型号每百万 token 输入 4 到 5 美元、输出 20 到 25 美元，旗舰档到 10 美元与 50 美元，整体站在第一梯队最贵的一侧。模型能力长期霸榜，价格也长期霸榜。
过去两年，企业买大模型看的是效果榜，谁强买谁。现在采购部门的逻辑换了：谁的 token 性价比高，谁留下。OpenAI 的 GPT 系列在今年连续降价，微软内部的 Copilot 从7月起把复杂查询默认路由到 OpenAI 模型，理由写得很直白，提升 token 投入的性价比。
好用但贵，这个组合在消费品里是品牌溢价，在企业采购里是第一个被优化的对象。尤其在裁员季，每个部门都要回答同一道题：这笔 AI 支出，省下来能养几个工程师？
03 两万亿估值的裂缝
把这些事放回 Anthropic 的时间表上，分量就清楚了。
按媒体报道，Anthropic 计划最快11月中旬完成 IPO，11月9日当周启动路演，争取感恩节前挂牌。10月22日前后又有消息说，公司最快当周公开提交 S-1 文件，留给它的窗口正在收窄。
财务底牌已经通过招股材料草案陆续露出。截至6月30日的六个月，公司营收约 163 亿美元，上年同期只有 11 亿美元；年化收入8月已超 650 亿美元，市场预计年底突破 1100 亿。
同时，上半年净亏损 807 亿美元，未来算力支出计划高达数千亿美元。
最刺眼的一行是客户结构：2025年，Anthropic 近四分之一的收入来自两家客户。媒体没有点名，但微软和 Meta 作为其最大两家企业客户，很难不让人对号入座。招股书的风险因素里还写了一句，最大的一批客户中，不少没有签长期合同，可能减少甚至停止支出。
现在这句话正在变成现实。两个最大的金主在同一个月份收缩内部用量，市场的反应是直接调预期：部分投资者已把 Anthropic 的合理估值从约 2 万亿美元下调到约 1.5 万亿。同期 AI 板块也在抛售，CoreWeave 两周跌去约三成。
Anthropic 的处境可以概括成一个错位。收入增速是真的，半年十几倍；客户的忠诚度却是有条件的，条件叫价格。它的增长引擎是少数巨型客户的海量消耗，而这些客户恰好是全球最有能力自研替代、也最精于算账的一批公司。大厂把 AI 能力收回内部，是一个已经开始、很难逆转的动作。
接下来三周有几个日子值得盯着：S-1 公开文件里的客户集中度会写得多细，路演中管理层怎么回答微软和 Meta 的问题，以及感恩节前能不能如愿挂牌。如果挂牌顺利，Claude 贵不贵是华尔街用真金白银投票的事；如果推迟，今天这两个金主踩下的刹车，就会被写进每一份推迟原因的分析里。
对国内的大模型公司，这件事的参照意义也很直接。靠少数大客户把收入冲高的路径， Anthropic 已经演示了它的天花板在哪。价格能不能接住能力，客户结构能不能摊开，这两个问题早晚要答，上市只是让答案变成了公开信息。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
♪ Man, ♪
On 24 October, within 45 minutes, Claude Code ' s two most prominent members resigned from office in X.
One of the core engineers of the Thariq Shihipar, Claude Code team, is associated with the development community, known as "the King of Interpretation", the Skills system and the context project. The other one was Anthony Morris, who created Claude Code's desktop and said his last day in Anthropic was October 23rd.
Neither of them said where to go next, and the farewell letter was decent. Shihipar also left a good message to the former family saying that his colleagues were the most talented group he had ever met.
This day, only three weeks away from anthropic planned IPO. The company ' s target valuation peaked at $2 trillion and is expected to update the global IPO record. The road show hasn't started yet. The two most famous faces on the product line are gone.
One, two gold owners and a brake.
Separation is just this line on the stage. Earlier on October 5, The Information reported another thing: Anthropic's two largest corporate clients, Microsoft and Meta, were compressing the use of Claude by their own employees.
The Microsoft executive estimated earlier this year that the annual expenditure on the intra-corporate use of Anthropic technology would be at least $1 billion. Now more than a third of the budget has been cut off and employees have been asked to use less Claude, instead GitHub Copilot and self-study models.
More subtle is quota. Microsoft Clouds and AI had more than 60,000 employees, with a budget ceiling of $100,000 per person per month for the AI model, and most of the scenes had been crushed to about $10,000 in the recent past, with 90% cut. Engineers complain that the model is no longer available.
Meta over there, the in-house employees of Claude Code dropped from about 60,000 earlier this year to about 30,000. Part of the cut was the fallout from the spring, when Meta cut off about 10% of the jobs. But the main reason for this is that Meta is working on the tools: the dedicated internal Meta Code users are over 30,000, and the externally tested Muse Code has over 6000 employees.
There's a number worth looking at. After the user cut off, Meta spent more than $105 million on Claude Code for the last 28 days. Half the people are gone, the bills aren't too thin.
Microsoft cut off in-house office and R & D expenditures, and the portion of Claude purchased for the Copilot function for external clients, which was expected to be at least $2 billion per year, is reported to be rising steadily. Similarly, Meta and Microsoft continue to sell Claude to external clients through their respective cloud platforms.
So the exact expression is that the gold owner didn't withdraw the order, Claude, and the gold owner just stopped letting his employees open.
It's a crime.
Why did you contract? Rewinding Microsoft's time line, the answer is a little bit harsh: Claude Code is great.
In December 2025, Microsoft introduced Claude Code to thousands of engineers in the experience and equipment sector. No one's promoting, no one's mobilizing, it's on its own, and in the first few months, token consumption just doubled.
The problem then becomes an arithmetical question. Each engineer's token costs between $500 and $2,000 a month, and a team is millions of dollars. For the first time, AI tools, like cloud resources, need to be budgeted, audited and capped.
Behind this is Claude's price position. According to official pricing in October 2026, the Opus main power model fed between $4 and $5 per million token, exported between $20 and $25 per million, and the flagship bay between $10 and $50, standing as a whole on the most expensive side of the first team. Model capabilities have long prevailed, as have prices.
Over the past two years, companies have been looking at LLM's results list, who's strong enough to buy. Now the logic of the procurement department is changed: who's token has a higher value and who stays. Open AI's GPT series has been declining continuously this year, and Microsoft's internal Copilot has been routing complex queries by default to the Open AI model since July for reasons that are very clear and raise the value of the token input.
Useful but expensive, this combination is a brand premium in consumer goods and the first to be optimized in enterprise procurement. Especially during the layoffs, each department has to answer the same question: how many engineers can be saved from this AI expenditure?
03 2 trillion-valued cracks
Put these back on the Anthropic schedule, and the weight is clear.
According to media reports, Anthropic plans to complete the IPO by mid-November as soon as possible, and launch the road show that week on 9 November to get the cards up front for Thanksgiving. There was further information around 22 October that the company had been able to submit the S-1 document publicly at the earliest opportunity during the week and that the window left to it was narrowing.
The financial base card has been released through the draft stock call material. By 30 June, companies had collected approximately $16.3 billion, compared with $1.1 billion in the same period of the previous year; annualized revenues had exceeded $65 billion in August and the market was expected to reach $110 billion by the end of the year.
At the same time, there was a net loss of $80.7 billion in the first half of the year, and future computing plans amounted to hundreds of billions of dollars.
The most striking line is the client structure: in 2025, nearly a quarter of Anthropic's income came from two clients. The media did not name names, but Microsoft and Meta, as their two largest business clients, were hard to avoid. The risk factor of the book also states that many of the largest clients do not have permanent contracts, which may reduce or even stop spending.
That sentence is now becoming a reality. The two largest gold owners contracted their internal use in the same month, and the market responded with a direct shift in expectations: some investors had reduced the reasonable valuation of Anthropic from approximately $2 trillion to about $1.5 trillion. In the same period, the AI plate was also sold, and CoreWeave fell about 30 percent for two weeks.
Anthropic's situation can be summed up as an error. The increase in income is real, more than a dozen times in six months; the client's loyalty is conditional on a price. Its growth engine is the mass consumption of a small number of giant clients, who happen to be the most capable and better off firms in the world. It's a start-up, hard-to-reverse move to take back the AI capability.
The next three weeks will be worth looking at: the focus of clients in the S-1 public document will be much more detailed, how the road management will answer Microsoft and Meta’s questions, and whether they will be able to put up the cards before Thanksgiving. If the cards are put up well, Claude's expensive vote on Wall Street in real money and silver; if they are postponed, the brakes that the two gold owners stepped on today will be included in every analysis of the reasons for the delay.
The reference to LLM in the country is also direct. Anthropic has shown where its ceiling is. The question of whether prices hold up capacity and whether the client structure can be spread out is one of the two questions that will be answered sooner or later.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.