What does a website that keeps itself up to date cost?

My portfolio site updates itself every morning. At 08:15 a script reads my Notion database, rebuilds the pages in Dutch and in English, checks the links and the social cards, and puts the result live. I do not look at it.

Today I worked out what that costs. On real data: my own AI conversation logs from the past 45 days, the publication log, the change history of the site and the hours I record per task.

The daily routine uses zero AI

The biggest surprise first. The routine that updates the site every day uses no AI at all. No tokens, no credits. It is ordinary code: Python scripts that fetch data, build pages and run checks. That code was once written with AI, but it has been running ever since for the price of my laptop's electricity. In ten weeks: 46 automatic publications without me looking at any of them.

I see that pattern everywhere in my system. The AI costs sit in the building. Whatever runs as code afterwards costs next to nothing.

Where the AI does work: the build sessions

The building itself happens in working sessions with Claude: new pages, quality gates, articles, bug fixes. In 45 days that came to 28 sessions around the site. Added up from the logs: a little under a thousand euros in tokens, if I were paying the official API rate.

I am not. I have a subscription of €90 a month, and the site accounts for about 25% of what I use it for. Allocated: about €22 a month. Add image generation (about €5) and a slice of server rent (about €3) and the complete AI upkeep of the site costs about €30 a month in real money.

The gap between those two numbers was the most striking figure of the whole measurement: my total AI use has an API value of about €2,700 a month, against €90 in subscription fees. Anyone working through a subscription instead of the API is looking at a factor of roughly 30.

This is the moment to try

That factor-30 gap also says something about the market. Two public figures make it concrete. Sam Altman, the head of OpenAI, wrote in early 2025 that his company loses money on its 200-dollar-a-month subscriptions: people use them far more than he had estimated. And OpenAI's annual figures that leaked in June 2026 show that in 2025 the company took in over 13 billion dollars and booked almost 21 billion in operating losses: 1.60 dollars spent for every dollar that came in.

To be fair: that loss comes mostly from research and marketing, not demonstrably from running my subscription. And Anthropic, where my subscription lives, does not publish figures like these. The direction is clear enough to build on, though: the computing power I rent is currently being fronted by someone else.

For anyone wondering whether to build something with AI, the conclusion is simple: economically, this is the moment. Right now you rent computing power well below its asking price. That window may close once the growth phase ends and subscriptions move towards cost price. It may also stay open: models get more efficient every year and open-source alternatives push the price down. But I would not wait for that; trying costs the least right now.

The real cost: my own time

Thirty euros a month is a rounding error. The biggest cost is not on any invoice: my own steering time. Around 20 hours a month of setting direction, judging output and correcting course, measured through the actual effort I record for every finished task.

That is the lesson I put to everyone who wants to calculate "AI costs": the tokens are not the problem. The question is how much of your time it takes to get the output right. Leave those hours out and every sum comes out misleadingly cheap.

And fully by hand?

The same output without AI: publishing by hand every working day, with translation and checks (10 to 20 hours a month), writing dozens of portfolio items in two languages (6 to 13 hours), the construction work on pages and scripts (45 to 75 hours), plus the articles (6 to 8 hours). Together, 70 to 115 hours a month.

Against my 20 hours that is a factor of 3 to 5, or 50 to 95 hours gained per month. Though the honest conclusion is sharper: by hand, this volume would simply not exist. The site would have stayed a static one-pager that I update twice a year.

Heavy at the front, light afterwards

The hours measured over the months show the investment curve. June and July were build months: dozens of my own hours, one session after another, the peak of the summer holiday. In August I finished only a handful of site tasks and my own site hours dropped to a few a month.

Every part of this system works that way. The effort sits almost entirely at the front: building a publication routine or a quality gate takes an evening or a weekend, after which it runs and asks for almost nothing, in time or in money. Anyone who looks only at the build phase sees an expensive, labour-intensive project. Anyone who looks a few months later sees a site that keeps itself up to date for €30 a month.

AI usage actually peaked in August, while my own hours fell. So the ratio is tipping: more and more machine work per hour of mine. Exactly what you want to see in a system that is nearing completion: the human shifts from building to steering.

Do the sums yourself

Anyone who wants to know this for their own situation can start today: Claude Code writes local conversation logs with the token usage per session (the `~/.claude/projects` folder), and an afternoon with a short counting script gives you the same overview as this article. Two rules while you do it: value your tokens at what you actually pay (subscription, not API rate), and put your own review hours in the sum as the biggest line.