What a website built by AI costs: a public ledger

What I mostly hear about AI and websites is the same two words: fast and cheap. I rarely see exact numbers. So here are mine. From 19/09/2026 on, every site I build or change with AI goes in here: the date, what was done, and what landed on the invoice.

The first two jobs were done on the same day. One site for a school, one for a scrap metal and recycling yard. Both new sites are design concepts on a preview address, not live websites. Neither organisation commissioned or endorsed them. Below is what each job cost, what that sum is made of, and what I left out of it.

Tool

Claude Code on the Max 20x plan. The amounts are what its session panel reports, at API prices

Model

Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5, in different proportions in each job

Date run

19/09/2026

Task

Rebuild two live websites from start to finish

Cost

$160.35 for both jobs together

What was measured

The bill in dollars, the time spent, the number of pages and the weight of the home page

Amounts are in US dollars, because that is how Anthropic bills.

The websites

Old site on the left, new one on the right. Drag the middle. All the new ones are samples, not live sites.

Why one website cost more than two

The cheaper bill bought two complete versions of one site, the bill twice as big bought a single smaller site. Here is why.

Cost per page is the bill divided by the number of pages. It is a rough measure: the pages are not alike.

1. The most expensive model did all the hands-on work itself, instead of only checking the work of a cheaper one.
2. It used about 100 helper runs, and each helper first had to read the whole job in, which is also billed.
3. That job also included research on the business, 27 web searches, not only building the site.

Full explanation with numbers

The two builds side by side

	Richmond International School	ENI Environment
Pages built	28	10
The bill	$52.62	$107.73
Cost per page	$1.88	$10.77
Who did the bulk of the work	Opus 5 subagents. Fable 5.1 only planned and reviewed.	Fable 5.1 itself, plus about 100 subagent runs, most of them also on Fable.

Cost per page is worked out from the bill and the page count. It is a rough measure: the pages differ in size and in how hard they were.

Three reasons

1. The expensive model did the hands-on work

   Fable 5.1 costs twice what Opus 5 does for the same number of tokens. On Richmond, Fable only planned and reviewed, and Opus 5 subagents wrote the pages. On ENI, Fable wrote them itself. It is the same as letting the most senior person in the room write every line instead of reviewing the work of others.

2. About 100 helpers, and each one had to read itself in

   Every subagent starts empty. The first thing it does is read the brief and the files, and that reading is written into the cache, so it does not have to be read again, and billed at that model's price. ENI ran up 5.4M cache-write tokens, Richmond 1.1M.

3. The ENI job was more than building

   Before writing any page, the model researched the business itself: 27 web searches. Then came the image pipeline, the per-page OG cards, the structured data and a build check script. Richmond needed no research, because the content inventory was handed over as a file.

Where the tokens go

In both builds the text the model actually wrote (the code and the copy) is a small part of the bill: about 100k and 186k output tokens. Most of the tokens are the model re-reading the conversation and the files at every step: 78.6M and 71M cache reads. Each of those is cheap, but there are hundreds of times more of them.

So the split of the tokens and the split of the bill look different: one output token costs far more than one cache read.

Tokens by type

From the session logs. They run a little past the session panel, so the figures are approximate. The bars show true proportions, which is why output and fresh input are barely visible slivers. That is the point.

Richmond International School

• Output: 100k
• Cache writes: 1.1M
• Cache reads: 78.6M
• Fresh input: under 1k

ENI Environment

• Output: 186k
• Cache writes: 5.4M
• Cache reads: 71M
• Fresh input: about 1k

Cost share by model

From the session panel. There are no dollar amounts per token type here, because the panel does not show them.

Richmond International School

• Opus 5 80 %
• Fable 5.1 16 %
• Sonnet 5 4 %

ENI Environment

• Fable 5.1 81 %
• Opus 5 18 %
• Haiku 4.5 1 %

What I do differently now: plan and review with the most expensive model, hand the building to the cheaper one, and give the helpers a short brief instead of the whole conversation.

Plain-words glossary

Model

The AI program itself. There are several, from cheaper to dearer, and they work at different levels.

Tokens

The way the text a model reads and writes is counted, roughly a piece of a word. This is what you pay for.

Cache

A holding place for text already read, so it does not have to be read again. Writing into it costs money, reading from it is cheap.

Subagent

A separately started model that does one part of the job. It starts from nothing, so it first has to read the job in.

API key

The pay-per-use route. The bill arrives for each job separately.

Subscription

A monthly fee. You pay the same amount, but you can only do so much work in a week.

What to do on Monday: if you let AI write code, look at which model is doing the hands-on work. Leave the planning and the checking to the most expensive one, and hand the writing to a cheaper one.

What these numbers leave out

Building the site is only part of the job

What AI makes here is a folder of files: the pages, the design, the text. Two things are still missing before it is a working website.

First, the site has to live somewhere and be looked after: hosting, the domain, the email that comes from the forms.

Second, someone has to decide how the owner will change a price or add a news item later. That is what a content management system (CMS) is for. These sites have none: today every change means going back to the AI or to a developer.

The old Richmond site runs on Wix, where hosting and an editor come with one monthly fee. Leave that, and both have to be solved separately.

• My time: the reading, the checking and the decisions. The minutes of active attention are written next to each entry.
• Knowing what to ask for and what a good site needs.
• The owner confirming facts: phone, hours, photo rights.
• Hosting and the domain.
• A content management system (CMS), so the owner can change the text without help.
• Content the business still has to supply.
• Upkeep: every later change gets its own row in the ledger.

A fair objection: these are not the prices of websites, they are the prices of a folder of files. None of these sites runs anywhere, none has a content management system, and nobody asked for them. Agreed, which is why what the numbers leave out has a section of its own.

What it would cost you

First, about me. I work on the Max 20x plan, which costs $200 a month, so I never got these amounts as a bill. They are what the same work would cost at API prices, as Claude Code itself counts it. For me it is not money, it is a share of the weekly limit.

Prices, plans and limits

API prices per million tokens

Model	Input	Cache write 5 min	Cache write 1 h	Cache read	Output
Fable 5.1	$10	$12.50	$20	$0.25	$50
Opus 5	$5	$6.25	$10	$0.50	$25
Sonnet 5	$2	$2.50	$4	$0.20	$10
Haiku 4.5	$1	$1.25	$2	$0.10	$5

platform.claude.com, read 19/09/2026.

Subscriptions

Pro$20 a month ($17 billed annually)Claude Code included. Fable models are not inside the plan limits: they run on usage credits, billed at standard API rates.

Max 5x$100 a monthClaude Code included. Fable 5.1 included, up to 50 % of the weekly usage limit, after that usage credits or another model.

Max 20x$200 a monthThe same as Max 5x, with higher limits.

claude.com/pricing and support.claude.com, read 19/09/2026.

How the limits work

A rolling 5-hour session limit plus a weekly limit, shared between Claude Code and the Claude apps. At the limit you wait for the reset or turn on usage credits, which are billed at standard API rates.

Anthropic's own figures

Anthropic puts Claude Code on the API at about $13 per developer per active day and $150 to $250 per developer per month, with 90 % of users under $30 per active day.

Anthropic does not publish how many dollars of API-equivalent usage a plan covers, so this page does not claim one.

Three routes to the same work

	API key	Pro	Max
What you pay for everything in this ledger	$160.35 to date. That is what lands on the invoice.	The Fable part would be billed at API rates anyway. In practice Pro means doing this with Opus and Sonnet only.	$0 extra, but it eats part of the week's limit. How much is not shown anywhere.
What limits you	Nothing except money.	The rolling 5-hour limit and the weekly limit. You wait them out.	The same limits, higher. Fable up to 50 % of the weekly limit.
Which models	All of them, at the prices in the table above.	The Claude Code models in the plan. Fable only on usage credits.	The same ones, plus Fable 5.1 up to 50 % of the weekly limit.
Who it suits	Someone who wants the exact cost of each job.	Someone working in smaller bursts who can wait for a reset.	Someone working daily who does not want to think about the bill.

Plain-words glossary

API key

The pay-per-use route. The bill arrives for each job separately.

Subscription

A monthly fee. You pay the same amount, but you can only do so much work in a week.

Over time

Every job gets its own row with a date and a price. The list grows with each new job.

Sources

• Claude API pricing: platform.claude.com/docs/en/about-claude/pricing (read 19/09/2026)
• Claude plans: claude.com/pricing (read 19/09/2026)
• Claude support: usage limits: support.claude.com (read 19/09/2026)
• Claude Code costs: code.claude.com/docs/en/costs (read 19/09/2026)

The bill for a website stopped being a mystery to me and became a row in a table. The interesting part is not that it is small. It is how easily one careless decision doubles it: who does the work, who only checks it, and how many times the same text gets read again.

I wrote about the limits I keep when I build with words separately: vibe coding with limits. What else changed over this summer is here.