Anthropic Says Its Own AI Use Is Now Worth More Than Its Payroll. What That Number Means, and What It Doesn’t

By Thorsten Meyer | November 1, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Somewhere in April 2026, a line crossed another line inside Anthropic. The company behind the Claude models says that in that month, the value of the AI its own staff used for one month was higher than what it paid those staff for the same month, leaving out stock awards. Business Insider picked the sentence out of Anthropic’s IPO prospectus on 1 November and called it a remarkable milestone. It is one. It is also a number built on a price tag, not on a bill.

A note before we start: Anthropic makes Claude, and Claude helped research and draft this article. That is a reason to read Anthropic’s numbers more critically here, not less. Everything below comes from the filing itself, from what Business Insider reported, or from my own arithmetic, and I say which is which.

Key takeaways

• The source is Anthropic’s own S-1. In the prospectus it filed with the US Securities and Exchange Commission on 26 October 2026, Anthropic writes that the value of its internal Claude usage “at list prices has grown by over 250 times from January 2026 to September 2026, and in April 2026, this monthly value first exceeded our monthly personnel expense, excluding stock-based compensation” (p. 185).
• “At list prices” is not cost. Anthropic priced its own usage at what customers pay. Its real cost is the compute behind that usage, and the filing does not break it out. As Business Insider put it, Anthropic “isn’t actually spending more cash on AI than salaries.”
• “Excluding stock-based compensation” matters a lot. Stock awards made up about 70% of Anthropic’s personnel expense in the first half of 2026 (own calculation from the filing). The bar the AI cleared is the cash part of pay: salaries, bonuses, benefits and payroll taxes.
• No dollar figure is given. Working back from the quarterly numbers, cash personnel cost ran at roughly $124 million a month in the first quarter and about $186 million a month in the second (own calculation). So April’s crossover sits somewhere in that range. That is an estimate, not a disclosure.
• The AI did not replace the staff. Anthropic’s headcount grew more than five times between the end of 2024 and September 2026, to about 5,300 people. The AI was added on top of a fast-growing team, not swapped in for one.
• The real story is the shape of the company: more than 90% of code written by Claude, revenue up 86-fold against headcount up five-fold, and a risk section that warns the company “may lose visibility” over processes its models increasingly run.

Products Worth Considering

DJI Osmo Mobile 8 Gimbal with Mic Mini Transmitter, Infinity Black

AmazonView Latest Price

LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor

AmazonView Latest Price

DJI Osmo Mobile 8 Standard Combo Gimbal Stabilizer for Phone

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

What the filing actually says

The sentence sits in the “Business” section of the S-1, under the heading “Claude is Accelerating Our Ability to Close the Loop.” Here it is in full:

“The value of our own internal Claude usage at list prices has grown by over 250 times from January 2026 to September 2026, and in April 2026, this monthly value first exceeded our monthly personnel expense, excluding stock-based compensation.”

The next sentence explains why Anthropic wants you to read it: “Because each version of Claude is used to help build the next and because Claude makes our teams more productive across the company, we believe the compute we devote to internal usage compounds these gains, continuously accelerating our pace of development and further strengthening our leadership position.”

Four things are worth pinning down, because headlines tend to blur them:

• The period. The 250-fold growth runs from January to September 2026. The crossover happened in one month, April 2026, and is described as the first time. The filing does not say whether it stayed above in every month after that, or by how much it is above today.
• The method. Internal usage is valued “at list prices,” meaning the prices Anthropic charges customers. The filing does not say which models, which price list, or whether any discounts large customers get were applied.
• The comparison. It is monthly usage value against monthly personnel expense without stock awards. Anthropic defines personnel expense as stock-based compensation, cash compensation “including base salaries and bonuses,” employee benefits and employer payroll taxes (p. 147). Take the stock away and what is left is the cash side of pay.
• The status. This is an operating figure from Anthropic’s own systems. It is not a line in the audited financial statements. The filing’s own risk section warns that several of its operating metrics “have not been independently verified by third parties” (p. 68).

Price tag versus bill: why “at list prices” changes the meaning

Imagine a bakery that eats its own bread. If it values every loaf at the shop price, the staff lunch looks expensive. What the bakery actually spent is flour, ovens and electricity. Anthropic’s number is the shop price.

The filing gives a hint of how big that gap can be. In the first half of 2026, Anthropic’s cost of revenue, which it says consists substantially of the compute used to serve customers, was $2.38 billion against revenue of $16.29 billion (Note 8, p. F-40). By my arithmetic, that is about 15 cents of serving cost per dollar customers paid. Internal usage is not in that line: Anthropic books the compute for “model development, research, and internal usage workloads” under research and development. So we cannot know the real cost of its own Claude use. But if it is anything like the cost of serving customers, the bill is a fraction of the list value.

That cuts both ways. It means the headline overstates what Anthropic spends. It also means Anthropic gets its AI co-workers at cost, while every customer pays the list price. No other company can copy that ratio on the same terms.

What the AI beat: the cash side of payroll

Stock awards are a huge part of pay at Anthropic. According to the filing, personnel expense in the first half of 2026 was $3.15 billion, of which $2.2 billion was stock-based compensation (p. 23). That is about 70%. Remove it, and the cash side of pay was roughly $0.95 billion for six months.

The quarterly table on page 158 lets us get closer to April:

• First quarter 2026: personnel expense $1,295 million, including $924 million in stock awards. Cash side: about $371 million, or roughly $124 million a month.
• Second quarter 2026: personnel expense $1,857 million, including about $1.3 billion in stock awards (the filing rounds this figure). Cash side: about $557 million, or roughly $186 million a month.

April is in the second quarter. So the monthly value of Anthropic’s own Claude use, priced at list, first passed something in the range of $124 million to $186 million. That is my estimate from the filing, not a number Anthropic gives. For scale: Anthropic’s revenue in the second quarter alone was $11.56 billion, and it spent $5.56 billion on compute and infrastructure in the same three months (p. 158).

The company this number describes

The payroll comparison is the catchy part. The surrounding paragraphs in the S-1 describe something more unusual: a company where the product does a large share of the company’s own work.

• Code. “Over 90% of the contributions to our codebase in September 2026 were written by Claude,” Anthropic writes. The average daily volume of code contributed per engineer in the third quarter of 2026 was “over 45 times” the 2024 average (p. 184). Volume of code is not the same as value of code, but the direction is clear.
• Hard problems. On “challenging, open-ended engineering problems” that used to need experienced human judgment, Claude’s success rate reached about 90% in September 2026, up about 50 percentage points in nine months. That is measured on Anthropic’s internal benchmarks (p. 184).
• Headcount versus revenue. From the end of 2024 to 30 September 2026, headcount grew more than five times, while revenue for the third quarter of 2026 was 86 times the revenue of the fourth quarter of 2024 (p. 184). With about 5,300 employees, trailing twelve-month revenue per employee was over $6.6 million (p. 196).

Anthropic’s reading is that “Claude allows employees to leverage a team of agents in their daily work, enabling a comparatively small team to take on work that would traditionally require a far larger one.” In its pitch to customers it calls these systems “Virtual Collaborators” and says it wants to extend them to knowledge workers globally, a group it puts at more than one billion people by 2028, citing IDC (pp. 8–9).

Here is the critical reading. An IPO prospectus is a sales document with legal liability attached. The internal-usage figure doubles as a product demo: look how much our own people use it. Revenue that grew 86-fold in seven quarters would make almost any staffing ratio look efficient, with or without AI. And the filing does not separate how much of the productivity comes from Claude and how much from a business that was growing explosively anyway.

Did the AI replace anyone at Anthropic?

No, not by the filing’s own numbers. Headcount grew more than five-fold between the end of 2024 and September 2026, and about 170% in the twelve months to 30 September 2026 alone (p. 137). Personnel expense in the first half of 2026 was up 486% on a year earlier, which Anthropic attributes to headcount growth (p. 150).

So the honest version of the headline is: at Anthropic, AI usage grew much faster than an already fast-growing payroll. The AI was added to people, not swapped in for them. That is also the pattern that makes “AI replaced X jobs” claims elsewhere worth checking carefully. I looked at how often companies blame AI for cuts the data doesn’t support in Is AI really behind the layoffs?.

What changes is the work. Anthropic says its engineers “increasingly direct and review” Claude’s work “rather than write code line by line.” That is a different job from the one most of them were hired for.

The warning in the same document

The most interesting sentences about internal AI use are not in the business section. They are in the risk factors.

Anthropic writes that because its models “write a large and growing share of our code and increasingly perform our research, including our safety research,” misaligned behavior in internal use “could undermine that work, introduce vulnerabilities into our systems or into the training of future models, or result in models operating outside of our intended controls.” And: “As we increasingly leverage Claude to accelerate progress within Anthropic, our organization may lose visibility across certain processes in a way that may become difficult for any individual or group of humans to understand and control.” It adds: “Our monitoring does not cover all of our internal and external deployments” (p. 49).

It is unusual for a company to say this about its own operations. It is also the right frame for the payroll number. When the AI line is bigger than the cash payroll, the important question is not only what the AI costs. It is who still understands what it does.

Anthropic also flags the labor side. It warns that public perception that its products “contribute to job displacement or economic disruption” could cause backlash “regardless of the actual impact” on employment (p. 55), and that governments may step in if job displacement “rapidly increases” (p. 70).

What this means for companies

• Treat AI as a cost line, priced at your price. Anthropic gets its own models at cost. You pay list or a negotiated rate. Before you compare your AI spend with your payroll, use your real invoices, not someone else’s list value.
• Measure outcomes, not tokens. A usage number that grows 250-fold says people use the tool. It does not say what they got for it. Amazon told its engineers the same thing in October. I covered that in Amazon tells engineers to stop tokenmaxxing.
• Budget for review, not only generation. If AI writes most of your code, human time moves to reviewing, testing and owning it. That time is a cost too, and it is where errors are caught.
• Keep visibility on purpose. Anthropic’s own risk factor is a checklist: know which agents run where, what they can touch, and which processes no single person understands anymore.
• Watch the tokens-as-pay idea. Business Insider argues that tech companies are starting to treat AI tokens as a fourth part of compensation, next to salary, bonus and stock. If you go there, write down who gets what budget and why, or it becomes a quiet new inequality inside teams.

What this means for workers

• The job is moving from doing to directing. At Anthropic, engineers increasingly “direct and review” rather than write line by line. Getting good at specifying, checking and judging AI output is the skill that transfers.
• An AI budget is not a raise. If your employer starts quoting the token budget it gives you as part of your package, remember that tokens are a tool you use for the company’s work. They don’t pay rent.
• Growth hides displacement, and slowdowns reveal it. At a company growing 86-fold, AI and hiring rise together. In a flat business, the same tools invite a different calculation. Ask how your employer plans to use productivity gains: more output, or fewer people. Anthropic’s own filing shows it expects that question.
• Your judgment is the scarce part. The work that still needed “experienced human judgment” is exactly what Anthropic measures its model against. That is where to stay strongest, and where to insist a human signs off.

Why you should be skeptical of this article, too

Anthropic is a central party here in two ways: it is the company being analysed, and its model helped write the analysis. I have tried to keep that from shaping the conclusions by sticking to the filing’s words, labelling my own arithmetic, and putting the critical points next to the impressive ones. The prospectus is also still a draft: the share count and price range are blank, and Anthropic can amend it before its listing. For context on the offering itself, see Anthropic could go public with a “junk” credit rating.

This article is information, not investment advice. It does not recommend buying, selling or holding any security.

What we don’t know yet

• The dollar value. The filing gives a growth multiple and a crossover month, not the value of internal usage.
• The real cost. The compute cost of internal use is inside research and development and not broken out.
• The price basis. Which models and which list prices were used, and whether discounts were applied.
• What happened after April. Whether the value stayed above cash payroll every month, and the ratio in September.
• How much is Claude and how much is growth. The productivity figures are internal, partly from Anthropic’s own benchmarks, and not independently verified.

The bottom line

Anthropic’s sentence is real and it is in a legal filing: in April 2026, the list-price value of the Claude its own staff used in a month passed what it paid them in cash that month. Read without the fine print, it sounds like AI now costs more than people. Read with it, it says something narrower and more interesting. Anthropic’s own AI usage is enormous and still growing, it is priced here at what customers pay rather than what it costs, it beat the cash part of pay while stock awards make up most of the total, and it came alongside a five-fold rise in headcount, not instead of one. The question it leaves every other company with is not “is our AI worth more than our payroll?” It is “do we still know what our AI is doing, and who is accountable for it?”

Sources

• Anthropic, PBC, Form S-1 Registration Statement, filed with the SEC on 26 October 2026 (internal usage at list prices p. 185; code, headcount and revenue p. 184; employees and revenue per employee p. 196; personnel expense definition p. 147 and growth p. 150; quarterly results p. 158; summary financial data p. 23; functional operating expenses, Note 8, p. F-40; risk factors pp. 49, 55, 68, 70; Virtual Collaborators pp. 8–9). sec.gov
• Anthropic, “Anthropic files S-1 with the SEC for proposed initial public offering,” 26 October 2026. anthropic.com
• SEC EDGAR filing index for Anthropic, PBC (checked 1 November 2026: no amendment to the S-1 filed). sec.gov
• Business Insider, Alistair Barr: “Anthropic’s internal AI use, valued at retail prices, now exceeds its employee pay,” 1 November 2026 (largely subscriber-only; cited for its framing and the “fourth component of compensation” argument). businessinsider.com

Own calculations in this article (cash personnel cost per quarter and month, stock share of personnel expense, cost of revenue as a share of revenue) use the S-1’s figures and are labelled as such. This article is for information only and is not investment, financial, legal or career advice.

Researched and drafted with AI assistance, including Anthropic’s Claude.