$109 of AI Got Me One-Sixth of a Poem. The Fix Cost $25.
I built a free web reader for public-domain books. It keeps the work in its original language. No translated edition, no English crutch.
Hover any word and you get its meaning in context. Not the dictionary's first guess. The sense that actually fits THAT line. Old texts reuse words in archaic ways, so "the" meaning isn't enough. You need the meaning here.
Click any phrase and you get a grammar explanation plus a full translation of the sentence.
One honest caveat. The word-level glossary is the weakest part. It picks the sense of a single word, and in older texts that is genuinely hard, because the language has shifted under it over centuries. So a hover meaning is sometimes wrong, especially in Dante. The grammar explanations and the sentence translations are far stronger, because there the AI gets the whole sentence, plus the one before it and the one after, as context instead of one word in isolation. More context, better answer. Trust the click more than the hover.
Live right now: the entire Divine Comedy in Italian. All 100 cantos. Inferno, Purgatorio, Paradiso. And Dostoevsky's Notes from Underground in Russian, all 21 chapters.
I built it for my own language learning. But the interesting part isn't the reader. It's the unit economics.
The trick that makes it cheap
Nothing runs when you open it.
Every word sense, every grammar note, every translation is precomputed offline ONCE. Then it's served as plain static JSON and pages. No model fires when a reader lands on a canto. It's fast, and it's almost free to host.
That's the whole method. Take a body of content, compute an AI layer over it one time, serve it cheap forever.
Beautiful in theory. Then I tried to pay for it.
What broke
I started on the flagship model. The expensive one. The good one.
$109 in, I checked the progress bar. I was about one-sixth of the way through the Divine Comedy.
One-sixth. For $109.
Do the arithmetic and the full poem alone was heading for $650 or more. And that's before the second book. My own up-front estimate was not a little off. It was off by a lot.
The silent killer was output cost.
This job doesn't summarize. It GENERATES. Long grammar notes and full sentence translations for thousands of sentences. And generated text bills far faster than the text you feed in. I was watching the input. The output was the part eating the money.
The fix was boring and it worked
Before scaling another dollar, I stopped.
I ran one chapter on a cheaper "mini" model. Then I put them side by side. Same chapter, both models. Quality next to quality. Cost next to cost. An honest head-to-head on a small, representative sample.
For this task, the mini held up. And it was 25 to 30 times cheaper.
So I finished the other five-sixths of the Comedy on the mini. Then I did the entire Notes from Underground on it too.
Total for everything after that first $109: about $25.
Read that again. The flagship bought me one-sixth of one book. $25 on the mini bought me the remaining five-sixths PLUS a whole second book in a different language.
The lesson
Calibrate cheap-versus-expensive on a small representative sample BEFORE you commit. Not after. The expensive model is not automatically the right one. For some jobs it's just expensive.
I caught this at one-sixth. Not at a $650 bill. That's the only reason this is a build-in-public post and not a confession.
One more thing that compounds. The reader keeps a shared glossary per language, so the vocabulary work carries between books. The first Russian book paid for the glossary. The second Russian book inherits it for almost nothing. The method gets cheaper the more you feed it.
The takeaway
The reader is a demo of a repeatable move: take a pile of content, precompute an AI layer over it once, serve it cheap forever. The skill that makes it pay isn't the model. It's measuring cost on a small sample before you scale.
The reader is live and free: read it here. The code is open source under the AGPL: on GitHub.
I build in public. Subscribe to the newsletter to follow what ships next.