What a website costs to build with AI

Cost per day

The spending is lumpy rather than steady. The two busiest days carry 58% of the whole total, and the dearest day cost 8 times what the quietest one did. The average of $53.87 a day is arithmetic rather than a rate you could plan against, because cost follows how much new ground a day covers.

Where the money goes

67% of the money goes on re-reading context the model has already seen, not on producing anything new. Writing the actual output is 13%. That ratio is the thing most cost estimates get wrong, because it is invisible from the outside. Long conversations are expensive for the same reason long meetings are.

What caching saved

98% of the tokens were context being read again rather than written fresh. Cached reading is billed at a tenth of the normal input rate, so the saving is larger than the build itself.

Caching saved about 6.0 times what the build itself cost.

The numbers

The same figures are published as JSON and CSV for anyone who wants to check or reuse them.

What that number actually means

Two different numbers answer the question of what a website costs to build with AI, and they are often confused for each other. The $377 above is what this work would have cost billed per token at published list rates. It is the only figure that can be measured request by request, which makes it the right one for comparing one week of work against another.

It is not what was paid. Both tools ran on monthly subscriptions, which charge the same flat fee whether you send one request or a thousand. On a project this size the subscription is the cheaper of the two, so the per token figure is better read as a measure of how much work happened than as a bill. Anyone pricing a build should budget for a monthly seat, not for $377.

Questions

How much does a website cost to build with AI?

This one has cost $377 so far in AI usage, across 1,270 requests over 7 working days. That splits into Claude Code at $191 and Codex at $186. The figure covers all of the design, content, code, SEO and release work done with AI assistance on this site.

Is that an actual bill?

No. Both tools ran on monthly subscriptions, which charge a flat fee rather than charging per token. This figure is what the same work would have cost at the published pay as you go rates. It is a fair way to compare one week against another, not an invoice.

Why is the token count so much larger than the cost suggests?

Because 98% of the tokens are context being read again rather than written fresh. Every time an assistant continues a conversation it re-reads everything that came before. Cached re-reading is billed at a tenth of the normal input rate, so a very large token count can still carry a modest cost.

What did caching actually save?

$2,261. Without caching the same work would have cost $2,638 rather than $377, roughly 6.0 times more. It is the single largest cost decision in the project, and it is on by default.

Which tools and models are counted?

Claude Opus 5 and GPT-5.6 Sol, used through Claude Code and Codex. Work done in the Claude or ChatGPT web apps is not counted, because those apps do not record token usage on this machine.

How accurate is this?

The token counts are exact. They are what each service reported for each request, not an estimate worked back from the text. Prices come from the official published rate cards for each provider, and are reviewed regularly so the totals follow any change. One figure is an assumption. OpenAI charges more past a context size it does not publish, so a boundary of 128,000 tokens is assumed. Claude has no such surcharge.

How often does this page update?

Every figure on this page is generated from the collected data, so the dates and totals move with it rather than being written by hand. It currently covers 31 July 2026 to 7 August 2026. Collection runs regularly because the source logs are deleted after 30 days, so anything not captured inside that window is lost for good.

Does the published data contain anything private?

No. Only dates, token counts, model names and totals are published. No prompt, no message content, no file name, no file path and no session identifier is read by the collector or written to the dataset. You can check that by opening the JSON or CSV.

Can I reuse these numbers?

Yes. The full dataset is published as JSON and CSV at stable addresses under a Creative Commons Attribution licence. Attribution to eduarddziak.com is all that is asked.