An interactive display that explains, in plain English, how AI language models read and write text — and what it costs. Built for a public exhibition at a library in Exeter.
Visitors tap one of five everyday questions and immediately see three things: how their words were chopped into tokens, the real answer the AI gave, and what that exchange actually cost.
It is designed to be used standing up, at a distance, on a phone or a laptop: large type, large tap targets, and very little to read.
Everything runs in the browser. There is no server, no API key and no network call while it is running. Every question, answer and token count is worked out in advance and shipped as a plain file, so the display keeps working even if the venue's Wi-Fi does not.
Open index.html in a browser, or serve the folder:
python3 -m http.server 8765
# then visit http://localhost:8765- Push this repository to GitHub.
- Go to Settings → Pages.
- Under Build and deployment, choose Deploy from a branch, then pick the
mainbranch and the/ (root)folder. - Save. The site appears at
https://<user>.github.io/<repo>/after a minute or two.
There is no build step and no workflow to configure — the repository root is the site.
Each question was genuinely sent to Claude Haiku 4.5. The answers on the page are its real replies, unedited.
| Tap | Your tokens | AI tokens | Cost |
|---|---|---|---|
| Boiling an egg | 14 | 246 | 0.098p |
| A quick message | 28 | 204 | 0.083p |
| A week of meals | 39 | 920 | 0.37p |
| A long letter | 618 | 394 | 0.20p |
| A cover letter | 666 | 472 | 0.24p |
They are arranged to make one point without needing a paragraph to explain it: A week of meals has almost the shortest question of the five, yet costs the most — because answers are charged at five times the rate of questions.
The examples mention no real or invented organisations. The letter simply comes from "your energy supplier"; the job advert is for "a community library service in Exeter".
A sixth and seventh example sit in their own section, after the million-times one, and they are the same question asked two ways: once with a page of rambling backstory and no limit on the answer, and once briefly with "give me 5 short bullet points" on the end.
| Your tokens | AI tokens | Cost | |
|---|---|---|---|
| The long way | 353 | 1,547 | 0.64p |
| The short way | 30 | 277 | 0.11p |
83% cheaper, for the same advice. Both answers are shown so visitors can check that claim for themselves. Asked a million times, that is the difference between £6,369 and £1,114 — and the scale picker lets you see exactly that.
Three tips sit underneath: say how long an answer you want, skip the backstory, and ask everything at once rather than a bit at a time — separate follow-up questions re-send the whole conversation, so combining them is cheaper.
Everything on the page is real, and the page says so in its own footnotes. In summary:
- The answers are genuine. Each was produced by Claude Haiku 4.5 and is shown unedited.
- The prices are the published ones for that model: $1.00 per million tokens read and $5.00 per million tokens written, converted at roughly £1 = $1.27.
- The token counts are very close, not exact. Each AI model divides text up slightly differently, and Claude's own method is not published. The counter used here agreed with Claude's to within about 3% on these examples.
- Real bills would be a little higher. Models also produce hidden "thinking" text that is charged for. The page counts only the words you can actually see.
- Prices and models change quickly. These figures were prepared in 2026.
The five questions live in build/prompts.json. To change them you will need the
Claude CLI, signed in:
$EDITOR build/prompts.json # 1. write your questions
rm build/responses/<id>.json # 2. clear the answers you changed
./build/run-haiku.sh # ask Claude Haiku 4.5 for new ones
node build/build-data.js # 3. rebuild the page's data
node build/verify.js # 4. check nothing brokePlace names, the currency conversion and the wording are all editable — see AGENTS.md for how the pieces fit together and which details are load-bearing.
Built as a public-engagement exhibit by CORE-AIx Lab.
If you write software with AI and want to see where your own tokens go, have a look at aGiTrack.