Anthropic goes public: Claude, today’s models, revenue and what comes next

Anthropic goes public: Claude, today’s models, revenue and what comes next

On 19 November, Anthropic, the company behind the Claude assistant, listed on Nasdaq under the ticker ANT. That same week I was working on this blog with Claude itself. So I wanted to look at the listing from two angles: as an entrepreneur who reads numbers, and as a user who pays a subscription.

The largest IPO in history

The figures below are those reported by CNBC on the day of the listing.

  • The IPO price was $83 per share, above the initial range of $70–78.
  • Anthropic sold 980 million shares and raised $81.3 billion. It is the largest IPO on record, ahead of SpaceX’s listing in June.
  • The stock opened at $111.50 and closed its first day at $97.17, 17% above the IPO price.
  • At the close, the company was worth roughly $1.8 trillion on a fully diluted basis, making it the seventh most valuable US company.

Amazon and Google together own about 28% of the company (14.7% and 13.7% respectively, according to the prospectus). Nvidia announced it would buy around $10 billion more in shares right after the listing.

The structure is unusual for a listed company. Anthropic is a public benefit corporation, a company required by its charter to pursue a public purpose as well as profit. A trust set up in 2023, the Long-Term Benefit Trust, can appoint and remove a majority of the board. The seven co-founders keep voting control through a separate class of shares. In short, whoever buys ANT buys a share of the profits, not control of the company.

Where the money comes from

What impressed me most is not the share price but how fast revenue is growing. According to the prospectus as reported by CNBC, Anthropic had revenue of $33.24 billion in the first nine months of 2026, roughly 13 times more than in the same period last year.

The company also reports an annualised revenue run rate of about $81 billion at the end of October, up from roughly $30 billion in April. A little caution is needed here. The figure is calculated by multiplying the last 28 days of revenue by 13, so one very strong month inflates it. The prospectus itself warns that growth could slow sharply.

Subscriptions or API: where the money really comes from

Anthropic splits its revenue into two categories: consumption, meaning payment based on how many tokens you use (mostly through the API), and subscriptions, meaning the monthly plans in the Claude app. The figures in the prospectus make it clear that the big money does not come from $20 subscriptions.

PeriodConsumption (pay per token)SubscriptionsTotal
2024$288M (75%)$98M (25%)$386M
2025$3.8B (83%)$789M (17%)$4.59B
First half of 2026$13.1B (80%)$3.2B (20%)$16.3B

Two clarifications. First, “consumption” is not only the raw API. It also includes metered use of Anthropic’s own apps, such as Claude Code in companies or the Enterprise plan, where you pay for a seat plus usage. That is why Claude Code revenue cannot be isolated. Second, almost half of revenue (48% in the first half of 2026) comes through AWS, Google Cloud and Microsoft Azure, which resell Claude models to their own customers.

Who pays

  • Large customers bring in about three quarters of the money. More than 11,900 customers each pay over $100,000 a year (up from about 950 a year earlier). Of these, more than 1,900 pay over $1 million, and about 250 over $10 million.
  • The five largest customers brought in about 25% of revenue in the first nine months of 2026. None exceeds 10%.
  • The rest, about 40 million customers, each pay less than $100,000 a year. This includes individual subscribers and small companies.
  • The huge gap comes from programming. Between July and September 2026, Enterprise users who did not use Claude Code spent less than $75 a month on average. The most active 5% of Claude Code users spent about 20 times as much.

What each type of customer pays

Prices on the Claude website in November 2026:

PlanFor whomPrice
Freeanyone who wants to try it$0, with limited usage
Proindividual user$20/month, or $17/month billed annually
Maxheavy user$100/month: 5x Pro usage, plus $100 in API credits
Ultravery heavy user$200/month: 10x Pro usage, plus $200 in API credits
Team, standard seatteams of 2–150 people$25/month, or $20/month billed annually
Team, premium seatteam members with heavy usage$125/month, or $100/month billed annually
Enterpriselarge companies$20 per seat per month, minimum 20 seats, plus usage at API rates
APIcompanies building productspay per million tokens, from $0.10 (Haiku) to $40 (Fable)

What it costs Anthropic

Costs per customer category are not published. The prospectus does not say how much a Pro subscriber costs compared with an API customer. Only the totals are visible:

  • Computing power is the biggest expense: $8.9 billion in the first half of 2026, or 55% of revenue. In 2025 it was 160% of revenue, meaning the company spent more on compute than it earned.
  • The roughly 85% gross margin often quoted in the press is calculated only from the direct cost of serving customers, about $2.4 billion in the first half of the year. Almost three quarters of compute costs sit under research and marketing, including the compute used by free users. With those included, some analysts estimate a real margin of about 65%.
  • Sales and partnerships cost $3.2 billion, 19% of revenue, mostly fees paid to the cloud providers that resell the models.

My view: for Anthropic, the $20 subscription is more of an entry point. The money comes from companies that put Claude to work for hours, especially on programming. A Max or Ultra subscriber who uses it heavily probably brings little profit, or even a loss, but that does not show up separately in the published figures.

Which Claude models exist today and what they are for

Anthropic now has four main models. All of them can read very long documents, up to one million tokens (roughly 700,000 words), and work with both text and images. The prices below are the developer prices per million input and output tokens.

  • Claude Fable 5.2 ($8 / $40). The most capable model, for hard reasoning and for agents working for a long time on a complex task.
  • Claude Opus 5.5 ($4 / $20). The model Anthropic recommends for most work: programming, analysis, long documents, agents.
  • Claude Sonnet 5.5 ($2 / $10). The balance between speed and intelligence. Good for applications used by many people at once.
  • Claude Haiku 5.5 (from $0.10 / $0.50). Fast and cheap, for high volumes: classification, data extraction, sorting messages.

For an entrepreneur who does not write code, the choice is simpler than it looks. A subscription to the Claude app gives you access to the large models without thinking about tokens. The API only matters when you want to build something: a customer assistant, an automated internal workflow, a tool that reads invoices. Then the rule is simple: start with the cheapest model that does the job and move up only if it is not enough.

How I use it

I use Claude every day. Not to write my thoughts for me, but for the part that used to eat my time: research, checking figures, structuring a text, technical work on the website. Rebuilding this blog, with dozens of articles, translations and SEO settings, would have taken months. With an AI agent working directly in the site, it took a few weeks.

What I learned is what I wrote about AI agents as the next step after chatbots: the value is not in the conversation but in the fact that the model can do things, not just answer. And one more, less comfortable lesson: AI is wrong with the same confidence with which it is right. Every important figure needs checking by a human. Including the ones in this article.

If you run a small company and do not know where to start, I wrote about 7 things a small business can automate with AI.

Outlook: bubble or infrastructure?

My view is that both answers are partly true.

The case for optimism. The revenue is real and is growing faster than at any software company I have seen in 30 years. Companies pay because AI saves them real money, especially in programming. And models are getting cheaper per unit: Opus 5.5 costs less today than Opus 4 did a year and a half ago, with far better performance.

The case for caution. A $1.8 trillion valuation assumes this growth continues for years. Computing costs are enormous, and competition is fierce: OpenAI, Google and Chinese open-source models are fighting for the same customers. And the run rate is an optimistic figure by design.

As an investor, I would treat ANT like any very new and very expensive stock: only with money you can afford to see halved. That is not investment advice, just the rule I apply to myself.

As an entrepreneur, my conclusion is clearer. Whatever the stock does, the tools already exist, are getting cheaper and more capable. Those who learn to work with them now will have an advantage, just like those who learned e-commerce in 2005. At AI Expo Europe we talked about what comes after the hype, and my answer stays the same: the work comes next.

Listing and revenue figures are those reported by CNBC and from Anthropic’s prospectus (the split by subscriptions, consumption, customers and costs via Kepler’s analysis of the prospectus); the revenue run rate is calculated by the company. Model and plan prices are those published by Anthropic in November 2026. This article is not investment advice.