I Pay $120 for Claude and Use $7,000 at API Prices. Why?

In this article you will learn:

• how much Claude I use compared to what I pay
• why Anthropic lets me, in the words of its own IPO filing
• when Claude's peak hours fall in Europe (CET), and whether they are still a thing

Quick answer

• I pay $120 a month for Claude. Used to their weekly limits, my plans give about $7,100 of it at API prices. That's roughly 59x what they cost.
• The API price is not the cost. At Anthropic's 85% gross margin, my $7,100 is maybe $1,100 of compute.
• Subscribers are a funnel. Anthropic's IPO filing calls subscribers like me "a top-of-funnel engine for our enterprise business". The money is in companies: 80% of the revenue is usage-based.
• Most subscribers never reach their limits. Like in a gym, the light users pay for the heavy ones.
• The GPUs are paid for anyway, but they don't sit idle. Anthropic was short of compute this year, and it raises limits when new data centers come online.

What I pay vs. what I use

In a recent post I told you how my AI bill grew in October. The Claude part of it is $120 a month now: a Max plan and a Pro plan. (At the end of October, Anthropic renamed its top plans: Max 5x is now just Max, and Max 20x is now Ultra.)

Until October 20, one $20 plan was enough - I didn't even use all of it, so those weeks say nothing about how much a plan can give. Then I kept hitting the limits, which is why I kept buying more.

How much can a plan give, then? Each account reports the share of its weekly limit used, so I matched those numbers against my logs, priced at API rates. My Max account hit 100% on November 2, and again a week later.

Max really is about 5x Pro, as Anthropic says. Used to their limits, my two plans give about $7,100 of Claude a month at API prices: about 59x what I pay.

So does Anthropic lose $7,000 on me every month? No - the API price is not what the compute costs.

In its IPO filing (the S-1), Anthropic reported $16.3 billion of revenue in the first half of 2026 and $2.4 billion of cost of revenue, "substantially all" of it compute and infrastructure. That's a gross margin of about 85%. If my tokens cost Anthropic the same 15% of their API price, my $7,100 is about $1,100 of compute.

Take it with a grain of salt. 85% is the average for the whole company, not the margin on a subscriber's tokens, and it leaves out the cost of training the models. But the order of magnitude holds: I cost Anthropic maybe nine times what I pay. So the question stands: why let me?

Subscribers are the funnel

Anthropic wants people like me to have fun with Claude at home, so that our companies pay API prices at work. The S-1 says it outright:

The same intelligence and capabilities we deliver to enterprises are also valuable to prosumers in their personal endeavors. These users provide strategic value as a top-of-funnel engine for our enterprise business — individuals who utilize Claude for their personal work and then champion it inside their organizations.

"Prosumers", in the filing's own words, are "individual users who subscribe to AI tools to perform knowledge work". That's me. And the money is on the other side of the funnel:

80% of the revenue is usage-based: companies paying per token, through the API or the cloud marketplaces. Subscriptions are 20%, and that includes Team and Enterprise seats, not only Pro and Max.

Three more details point the same way:

• Free users are a marketing cost. The compute "associated with providing free access" is booked under sales and marketing, not as cost of revenue.
• Limits are a ramp, not a wall. Hitting one is, in the filing's words, "a seamless bridge from subscription to incremental pay-for-intelligence revenue": you move to a higher tier or buy more. That's my October: one Pro, a second Pro, then an upgrade to Max. In six days.
• API credits in Max and Ultra. Since the rename, both plans come with credits "to build on the Claude Platform", worth as much as the plan itself: $100 or $200 a month. It's an invitation to the side of the business where the money is.

Anthropic even lists the risk: "individual adoption of our products may not translate into organizational adoption or expansion". It's a bet, and a deliberate one.

The GPUs are paid for anyway

Data center chips cost money whether they're busy or not. When demand is low, can the spare compute go to subscribers at almost no cost? And since I live in Poland, do I get the quiet hours, while the US hits the limits?

What says yes:

• The compute is a fixed cost. Anthropic's long-term compute contracts "obligate us to pay for resources regardless of actual utilization". The deals signed or extended in 2026 alone add about $417 billion of future payments.
• Load does follow the clock. In March, Anthropic doubled the 5-hour limits outside peak hours for two weeks. Then it made session limits drain faster during peak hours (weekdays, 5-11 AM Pacific), saying "~7 percent of users will hit session limits they wouldn't have before".

What says no:

• Anthropic was short of compute, not long. From the S-1: "In the past, we have significantly underestimated our compute requirements, which left us capacity constrained".
• Quiet hours don't sit idle. "We continuously direct compute to its highest-value use case, balancing the need to advance research and model development with the need to serve our customers". A quiet hour can train or test the next model.
• Generosity follows new capacity. On May 6, Anthropic announced it would use all of SpaceX's Colossus 1 data center (over 300 MW, over 220,000 NVIDIA GPUs). The same day, it doubled Claude Code's 5-hour limits and removed the peak-hour reduction for Pro and Max.

Claude's peak hours in Europe (CET)

Claude's peak hours are weekdays 14:00-20:00 Central European Time (13:00-19:00 in the UK). Our morning is off-peak, and the peak starts after lunch.

Did I get the quiet hours? I checked my own logs:

Half of them. My busiest hours are 10:00-14:00, while America is asleep, but a third of my usage still lands in the US peak 😅

Are peak hours still a thing? Not since May, at least for Claude Code on Pro and Max: the reduction is gone.

So it's half true. The compute is paid for in advance, and Anthropic does tune limits to the clock. But it isn't sitting on spare GPUs, and my time zone isn't what makes my plan cheap: a third of my usage lands in the peak anyway, and the clock no longer changes the limits.

Max 20x is now Ultra

Back to the rename from the end of October, the latest move on the limits. The prices stayed the same:

Ultra dropped the 5-hour limit (for verified accounts), but its weekly usage is now described as "10x Pro" on Claude's official pricing page. The old name said 20x, and the announcement didn't mention the change.

Were the 5.5 models rushed out for the IPO?

A company about to go public wants every quarter to look good, and generous subscriptions help the growth numbers. Did it rush its models for the same reason? The whole 5.5 family came out in the weeks before the public S-1 of October 26:

Three models in 15 days looks like dressing up for investors, and the S-1 itself credits the growth to "successive model and product releases". But the timeline shows the same steady rhythm all year, the draft S-1 went in back in June, and the releases didn't stop after the filing: Fable 5.2 followed on November 9. I'd call it not proven. If the IPO shaped anything, it's the pace of the whole year, not one family.

So why does Anthropic let me do it?

Put together:

1. I'm cheap marketing. I get hooked at home and bring Claude to work.
2. My $7,100 is a list price. The compute behind it is maybe a seventh of that, it's paid for in advance anyway, and it gets cheaper fast. In the second quarter of 2026, Anthropic's revenue grew 1,368% year over year, while its compute and infrastructure costs grew 300%. And with Opus 5.5, it cut customers' cost per task by about 40% compared to Opus 5, while keeping its margins.
3. Most subscribers never reach their limits. When Anthropic announced weekly limits in July 2025, it said they would affect "less than 5% of subscribers". It works like a gym: those who rarely come pay for those who come every day. My first $20 plan was one of the light ones - until it wasn't.
4. Some subscribers pay with data, too. Since August 2025, Free, Pro and Max users choose whether Anthropic can use their chats and coding sessions to improve Claude; business plans and the API are excluded. For a company training coding models, that's valuable material.
5. Limits are the valve. They tightened when compute was short and doubled when new capacity arrived. The generosity follows the compute.
6. The money is in companies. According to the S-1, more than 11,900 customers pay Anthropic over $100,000 a year.

Final thoughts

• Limits moved both ways this year. I wouldn't treat today's generosity as a promise - especially once the company has to answer to public shareholders.
• Know your multiplier. Mine is about 59x, and I only know it because I measure it.

I hope this was a meaningful read - if so, let me know in the comments. And if you track your own usage at API prices, I'm curious what your multiplier is :)

What's next

Back to the hosting series: the next episode explains the four rendering modes - SSG, SSR, CSR and ISR. See you there!

FAQ