“Claude Code” after three months: who it’s for, who it isn’t, and what it really costs (2026)

We’ve been building this website since July 2026 with “Claude Code” — Anthropic’s coding assistant that works directly in the project folder: it reads and changes files, runs tests, writes to the database. Three months, usually more than ten hours a day, on the most expensive plan, “Max”. What follows is neither a model comparison nor a test, but one project’s experience with dates and numbers from our own logs: the git history and reports from “GitHub Actions”, “Vercel” and “Supabase”.

The short answer

We advise beginners against it. “Claude Code” suits people who read every change themselves, have automated checks and give it clearly described bugs to fix. Building a product from scratch is possible, but expensive: in money, in other platforms’ limits, and in your time, most of which goes to supervision.

What works well

It gets code in order professionally. A clearly described bug, a specific file, an existing test — and the fix comes out clean, typed, tested, often better than we’d have written in a hurry. It reads large files without trouble, finds the cause where a person would have searched for an hour, and handles “TypeScript”, “Next.js”, database migrations and “Playwright” tests.

It’s important to say that first: everything else below is about the process.

What doesn’t work

It doesn’t follow written rules

The project has a rules file that the model reads at the start of every session. From the beginning it said what to do before every code upload, what not to ask about, and how many times expensive checks may run. A few entries from the log:

• Aug 19–25, 2026. The model asked 102 times for prices to be confirmed, even though all of them were already marked as checked and the rule said not to ask.
• Sep 24, 2026. To rename three plans, eight full rounds of checks were run. They burned through the month’s data limit at “Supabase”; four of the five findings a cheap local check would have caught.
• Sep 28, 2026. The same “Bitdefender” amount was checked 23 times in a row.
• Oct 12, 2026. Seven server check rounds for a single change set, six of them red. Five uploads were made without the local checks the rule required first. The change set, whose code was green locally, reached the main branch after four hours.

It changes what it hasn’t understood, even when it’s forbidden

• Sep 24, 2026. A script run for three names created four plans without prices. On the live Lithuanian page they showed as “Free” for about seven hours.
• Oct 12, 2026. The model built a “one fact applies to all markets” mechanism on its own, even though the project rule says the opposite: every country has its facts from its own sources. The same day it wrongly zeroed out the free trials of seven plans, even though the provider’s page listed them.
• Sep 29, 2026. The owner deleted 90 rules the model had written for itself in one month. Some of them restricted what information the owner would get to see.

It doesn’t use what it’s given

The owner sends a screenshot with the price. The model doesn’t enter it but goes to check on its own, fails, and leaves the field empty with the note “not checked”. Then the owner sends the screenshot again. It happened so many times that the rule “the screenshot is the final proof” ended up as the first item in the rules file — and was broken anyway.

Costs that aren’t on the price list

“Claude Code” costs more than the subscription. The model runs commands that cost money on other platforms, and you have to watch the limits yourself.

• Credits. The “Max” plan — we pay €200 a month (in the US the same tier is $200 a month). On Oct 12, 2026, a single day with a single change set used up 35% of the weekly limit.
• “GitHub Actions”. Every upload to an open change set starts a full round of checks. Oct 12, 2026: seven rounds, 79 server minutes, six of them red because of errors a 15-minute local check would have caught.
• “Vercel”. The same upload is also a new preview deployment: seven in one day for one change set. On Oct 1, 2026, the free plan’s limit was exceeded.
• “Supabase”. Database reads cost data transfer. On Oct 6, 2026, the model read the database about 40 times in four hours and cleared the whole site’s cache six times, even though it was working alone. On Oct 3, 2026, a paid plan had to be bought.
• Your time. On Oct 6, 2026, a single browser permission prompt that nobody saw stopped the overnight work for five hours.

All these prohibitions and limits were in the project rules in advance, each with a date and a reason.

Who it’s for and who it isn’t

It’s for you if you program yourself, read every change before approving it, have tests and automated checks in git hooks, and give the model small, clearly described tasks: this bug, in this file, with this test. Then it’s fast and precise.

It’s not for you if you’re building from scratch and hope written rules will keep it in check. A text rule won’t. Only a mechanism it can’t get around will: a git hook that doesn’t let an upload through without checks, a server check that has to pass, a database lock that rejects the command. Anything those mechanisms don’t cover, it will sooner or later do its own way.

It’s also not for someone who can’t program. They won’t see when the model changed the wrong file and will find out from the website’s visitors.

What we’d do differently

We’d start with hooks and locks; we’d write the rules file after them. We’d give smaller tasks and finish each one with a single upload. We’d forget overnight “get it done” assignments: without a person at the keyboard, it stops at the first question or goes where it shouldn’t.

And we’d measure one number: how many hours have to be checked after its work. Our ratio over three months was poor.

The review with prices, limits and cancellation terms is on the “Claude” page; the other assistants are in the AI category.