AI News

10 Things in Anthropic's Prospectus: Price Cuts Drive Up Revenue, Coding Dominates Paid Usage

HuxiuUpdated Nov 11 13:16

AI summary

The author Kuaidao Qingyi reviewed Anthropic's 335-page prospectus and pointed out that its revenue grew from $41 million in 2023 to $16.286 billion in the first half of 2026, and that after price cuts, usage growth far exceeded the price decline. The prospectus shows that in 2026Q3, about 80% of revenue came from new models released in the past 6 months, the net revenue retention rate of existing customers was about 900%, and as of September 2026, more than 90% of internal code commits were completed by Claude. The author also mentioned that Anthropic does not do image and video generation, and bans violating customers that contribute 5% to 7% of revenue.

Why read this

AI score 82

The author breaks down the revenue, retention, and organizational data in the prospectus one by one, which can be used to understand the real structure of paid AI coding and Anthropic's trade-offs.

Topics

Tags

Body · AI translation

[[**Anthropic's prospectus shows that AI price cuts actually drove revenue higher, with coding dominating paid usage, existing customers' annual spending reaching up to 9x; internal AI value has exceeded labor costs, and the company adheres to safety-first governance.**

**Key points**

**1. Price cuts drove up total usage and revenue**

Revenue was $41 million in 2023, $4.585 billion in 2025, and $16.286 billion in the first half of 2026; quarter-over-quarter growth rose to 144%, with usage growth far outpacing price cuts.

**2. Model shelf life is about 6 months, and coding is the core paid scenario**

In 2026Q3, about 80% of revenue came from new models released in the past 6 months; spending fell about 20% during the New Year code freeze period, and the top 5% of Claude Code users spent 20 times as much as ordinary enterprise users.

**3. Existing customer retention is 900%, and large-enterprise budgets are far from being unlocked**

Net revenue retention is about 900%; Eli Lilly expanded from 10,000 researchers to 50,000 employees within a year; more than half of the Fortune 500 are already customers, but average spending is less than 0.1% of their total spending.

**4. Internal AI productivity has exceeded some labor costs**

As of September 2026, more than 90% of code commits were completed by Claude, and the average daily code volume per engineer was 45 times that of 2024; in April 2026, the converted value of Claude usage exceeded labor costs for the first time.

**5. Safety trade-offs and governance lock in the company's direction**

It does not do image/video, focusing on coding, agents, and safety; it bans violating customers that contribute 5%-7% of revenue; the founders' letter of persuasion, and the PBC safeguards its mission with 50.1% voting rights and a trust. 2026-10-28 09:14Key points from Anthropic's prospectus: 10 things about the AI business and organizationKuaidao QingyiQuick overview

This article is from the WeChat public account: Kuaidao Qingyi, author: Kuaidao Qingyi

Anthropic released its prospectus in the past two days, with 335 pages in the main text.

When I first opened it, I also spent a long time mentally preparing myself. After all, a 300-plus-page all-English document is daunting just to think about. I originally expected it to be filled with all kinds of shocking but boring numbers.

But after opening it, I still froze for a moment.

Because the first few pages of the prospectus are not data tables, nor revenue growth curves, but a set of illustrations: petri dishes, biological structure diagrams, star charts, and hand-drawn scientific diagrams, with the texture of an old scientific journal.

You should know that this company may set the record for the largest tech IPO in U.S. stock market history.

Yet what it puts at the beginning is petri dishes and star charts. These pages are actually a letter jointly written by seven founders. I will talk about the content of the letter at the end, because it is a very thought-provoking part of these 335 pages.

Today I will not talk about stocks, valuation, or who can make how much money in this IPO.

I will just pick 10 things from this prospectus and talk about the company's products, organization, and some key judgments. At the end, I will also share my own views.

First thing: the cheaper the model, the bigger the total bill.

Anthropic has been continuously cutting model prices over the past year or so. Opus 4.5, released in November 2025, is more capable and cheaper. This September, Opus 5.5 cut the cost per task by another about 40%, Fable 5.1 by about 25%, and the cost of complex coding tasks by up to 45%.

Normally, if things are sold cheaper, total revenue should shrink accordingly. But Anthropic's revenue curve is completely reversed.

Full-year revenue was $41 million in 2023, $386 million in 2024, $4.585 billion in 2025, and already $16.286 billion in the first half of 2026 alone.

Quarter-over-quarter growth also jumped from 48% in 2025Q4 to 132% in 2026Q1, and then to 144% in 2026Q2.

Unit prices were pushed down, but total volume exploded instead, and usage growth far exceeded the price decline.

There is a technical term for this in economics: the Jevons paradox. The most classic example is that after steam engines became more coal-efficient, total coal use actually went up.

Second thing: the shelf life of a model is only 6 months.

There is a data point in the prospectus that is easy to overlook: about 80% of 2026Q3 revenue came from new model versions released in the past 6 months.

In other words, models from half a year ago now contribute only 20%.

For Anthropic, this is both good news and bad news. The good news is that every time a new model is launched, customers upgrade along with it, and migration is fast. The bad news is that you cannot stop. Even if the release rhythm slows by one quarter, revenue may immediately feel the pressure.

In the first 9 months, 250 new model versions, platform features, and applications were launched, averaging one per day. At first, I thought large models were just a money-printing machine, continuously collecting money.

But now it seems that it is not an asset that pays off once and for all, but more like a crop harvested once a quarter. Only what is planted this season yields grain, and the surplus from the previous season is quickly eaten up.

Moreover, this is only Anthropic's own model releases. If you add OpenAI, Google, Grok, and DeepSeek, then for us users, choosing will indeed become a very exhausting task.

Third thing: when programmers take a holiday, revenue falls.

The prospectus mentions that in the two weeks before New Year's Day 2026, when many companies entered a code freeze period, Claude customers' total spending was about 20% lower than in the preceding two weeks.

You see, one number exposes the most core use case: Claude's biggest current revenue driver is programming.

Another set of data points in the same direction. Among Claude Enterprise users, those who do not use Claude Code spend less than $75 per month on average, while the top 5% of users who use Claude Code spend about 20 times that amount.

Many people use Claude perhaps for writing, drawing, making PPTs, or chatting. These use cases all exist, but the ones that truly generate large-scale payments are still programmers. From Monday to Friday at work, they use Claude to write code, debug, and do code reviews; once the weekend and holidays arrive, they put it down.

Of course, Claude has also noticed this problem. For example, in the past two months, they focused on building mobile remote control functionality. It is equivalent to being able to open your phone on the weekend and let the company computer continue writing code.

I have to say, this is really intense competition.

Fourth thing: existing customers spend 9 times as much in a year.

The prospectus mentions that, excluding Claude Pro and Max subscription customers, Anthropic's net revenue retention rate is about 900%.

This number deserves emphasis. Simply put, this net revenue retention rate means that the same group of existing customers spent, in this quarter, the equivalent of 9 times what they spent in the same quarter a year earlier.

You should know that in the traditional SaaS industry, reaching 120% is already considered good, and 150% is top-tier—meaning that users who have already signed up can contribute a bit more revenue in the second year.

Yet Anthropic's number is actually 9 times. I read it several times and could hardly believe my eyes, because this is almost another species.

Why is this happening? Because enterprise application scenarios are spreading. For example, the pharmaceutical company Eli Lilly is a typical case. In July 2025, Eli Lilly opened Claude Enterprise, the enterprise edition, to about 10,000 researchers.

In less than a year, by April 2026, its use had expanded to all more than 50,000 employees, covering multiple departments such as legal, finance, technology, and oncology.

An enterprise customer may initially pilot in just one department, and as they use it, it spreads to other departments. The models are also getting stronger; what could not be done half a year ago can now be done, and what can be done becomes new budget.

So existing customers will become increasingly tightly bound.

Fifth thing: more than 50% of the Fortune 500 are customers, but their spending is still less than a fraction of their budgets.

The prospectus states that all of the Fortune 10 and more than half of the Fortune 500 are already Anthropic customers.

Let me explain here: the Fortune 500 is the list of the 500 largest companies in the United States ranked annually by revenue by Fortune magazine. In other words, a large number of the most powerful companies have already become its customers.

By industry, all of the 10 largest banks are on the list, 9 of the top 10 pharmaceutical companies, and 9 of the top 10 chip companies.

Sounds impressive, right? But immediately after that there is another number: the money these Fortune 500 customers spend on Claude averages less than 0.1% of their total spending. Even the top 10 deepest users spend less than 1%.

One-thousandth—this is still a trial state. So what can be seen from this number? It is that large enterprises have already bought tickets and entered the venue, but most of their budgets are still sitting idle.

Once AI moves from pilot to standard configuration, even if the spending share rises from one-thousandth to one percent, for Anthropic that would mean tenfold growth space.

Sixth thing: 90% of its own code is written by Claude.

Next, let me tell you about some organizational changes seen in the prospectus. There is a sentence in the prospectus that is easy to gloss over as marketing talk:

As of September 2026, more than 90% of code commits in Anthropic's internal codebase were completed by Claude. The average amount of code contributed per engineer per day is more than 45 times that of 2024.

With about 5,300 employees, their biggest "colleague" is the model they built themselves.

They sell coding capability externally, and internally they also hand code over to their own model to write. Whether a company truly believes in its own product—this one item is enough to tell.

More than 90% of code commits completed by Claude—I completely believe this number. Because our product and research team now works on code the same way.

And it is not like before, specifying exactly which lines of code to write, but directly waking up the agent in the Feishu group and letting it write code and run experiments. People are mainly responsible for reviewing the final results.

At least in the past two months, the way our internal product and research team works has also been completely refreshed.

From January to September 2026, Anthropic's internal usage of Claude, converted to value at public list prices, grew more than 250 times. And in April 2026, the converted value of Claude's usage that month exceeded the company's labor costs for that month, excluding stock-based compensation, for the first time.

This number calculates: if the work Claude does were sold on the market, how much would it be worth.

To put it another way, 5,300 employees collect a salary every month, and Claude is also doing work every month. Starting in April, the work Claude does, converted to market value, was worth more than all the employees' salaries combined.

This is a milestone in organizational form: the most "expensive" labor inside a company is no longer just people.

Another change occurred over a longer time span: from the end of 2024 to the end of September 2026, the company's employees grew from about 940 to about 5,300 today, but quarterly revenue grew 86-fold.

Let's look at another number: revenue per employee over the trailing 12 months ending September 2026 exceeded $6.6 million, while a year earlier that number was only $1.4 million.

A term that often comes up in the AI era is "cost reduction and efficiency improvement." However, from Anthropic's example, we can see that companies should go all out for efficiency improvement, rather than using cost reduction as a pretext for disguised layoffs.

The eighth thing: deliberately not doing image generation, and not doing video either.

The prospectus specifically mentions two things Anthropic chose not to do. OpenAI is competing on images, Google is competing on video, and Meta is competing on multimodality. Anthropic is not competing in any of these tracks.

The prospectus gives three reasons: First, prioritize the direction with the greatest economic impact. Second, coding and agent scenarios are themselves testing grounds for safety research; safety issues will be genuinely exposed in these scenarios.

Third, and the one I find most interesting: code output can be verified. Code can run tests, right and wrong are relatively clear, and model training thus has a clear feedback signal. Whether an image looks good or whether a video is right is hard to judge in the same way.

The compute saved by not doing these things was all invested in coding, agent capabilities, and safety research.

For me, this may be the most interesting of the 10 insights. What a company chooses to do roughly reveals its capabilities; what it chooses not to do reveals its judgment.

The ninth thing: it banned violating customers who contributed 5% to 7% of revenue.

This point actually follows the same logic as the previous one. Anthropic disclosed in the prospectus that in the first nine months of 2026, about 5% to 7% of revenue came from customers later found to have violated usage policies or commercial terms.

This proportion is actually not small. And the prospectus states plainly how this works: the money is collected first, and accounts are only restricted or banned after violations are discovered. Because detection is slow, this portion has always been part of revenue.

In August and September, the company specifically strengthened detection and enforcement, and expects this portion of revenue to decline later, and it may also slow overall growth.

More interestingly, the prospectus also admits that banned accounts will register again with new accounts. After a ban, the money may still come back. By the way, everyone in the industry knows that Singapore is one of Claude's largest markets by traffic. Those who know, know.

In this sense, many domestic users probably have complicated feelings about this company: from the perspective of values and corporate culture, everyone deeply dislikes some of this company's practices; but in terms of model capability and ability to complete tasks, it is hard to resist.

So this is not a clean sweep after a blanket ban, but a tug-of-war that will continue in the future.

The tenth thing: the seven founders wrote a dissuasion letter at the beginning of the prospectus.

Back to the letter mentioned at the very beginning.

In the opening of most companies' prospectuses, the CEO writes a letter to investors. The formula is basically the same: we are great, we will be even greater, welcome to join our journey.

Anthropic's seven founders jointly wrote a letter that is completely the opposite.

They even say that investing in this company involves enormous uncertainty. If you cannot bear this uncertainty, please consider carefully.

A company sprinting toward the largest tech IPO in U.S. stock market history tells investors at the beginning of its prospectus: please consider carefully.

In this letter, the founders themselves use a term: anti-sell. That is, a "dissuasion letter," first laying out the company's trade-offs clearly and letting those willing to accept these trade-offs stay.

One more thing worth noting: Anthropic's full company name is Anthropic PBC. What does PBC mean? It refers to a public benefit corporation. It has two layers of locks in its corporate structure.

First layer: the seven founders hold 50.1% of voting rights through Class F shares, ensuring that after capital comes in, the company's direction will not be voted away. Even if investors have money, they cannot easily change their direction.

Second layer: an institution called the Long-Term Benefit Trust controls a majority of board seats and urges the company to balance the public interest. The trust has five trustee seats, though one was vacant at the time of disclosure.

Trustees cannot directly buy or hold the company's securities; they receive fixed cash compensation, and the compensation is not linked to company profits.

Together, the two layers are meant to ensure that the company's mission cannot be overturned by the shareholders' meeting, nor kidnapped by the founders' own short-term interests.

Looking back at those first few pages of petri dishes and star charts, you probably understand.

Over time, many people will surely say they are being pretentious. For example, they repeatedly emphasize safety and values; for example, they adopt the strictest blocking measures against domestic users; and for example, because they refused to let the U.S. military use them, they themselves were put on a blacklist.

At first, as a pragmatist, I always felt that they had a lot of performative elements. But in the past two years, whenever I go to Silicon Valley, I almost always seek out several Anthropic employees for in-depth conversations. Because I was curious: how can these Chinese people survive in such an "anti-China" company.

After talking with them, I found that these few people are extremely similar, and they all very much believe in their company's narrative. Including when I went in August this year, a friend gave me two copies of Anthropic's internal constitution booklet.

After flipping through them on and off, my biggest impression is that, on this matter, they are relatively consistent internally and externally, at least not like some companies that say one thing and do another.

From the booklet on my desk to this freshly released prospectus, this company has been emphasizing one thing: we are doing something very important and also very dangerous, and we ourselves are not sure how it will end.

Okay, these are some of the interesting details I saw in this prospectus.

Of course, for a company that the whole world is paying extraordinary attention to, in a prospectus of more than 300 pages, there must still be many interesting things I have not uncovered.

Original text of Anthropic's prospectus:

Has published 112 articles on Huxiu. This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. If there is any copyright issue, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]

Source

Huxiu