The goal was simple: analyse one company 90% cheaper than in my own ChatGPT or Claude — at the same quality.
Why it's worth solving for me: whether or not the price of tokens goes up, burning them for nothing makes no sense. And if analysing one company costs you over $10 today, hidden in a monthly fee, picture the bill when you analyse hundreds of companies a month.
I read text, not images
I pull the financial statements from the collection of deeds in the Czech commercial register. They're PDFs, and I used to send every page to the model as an image. An image is expensive — one set of statements swallowed 100 to 270 thousand tokens (tokens are what you pay for with AI).
Now I read the text layer the PDF already contains. The output is identical, just cheaper. For Barrandov Studio a.s., one filing dropped from 150 to 37 thousand tokens — four times less.
It isn't always that big a jump. For scanned filings, where some pages still have to be read as images, the saving is smaller — for MSV Metal Studénka, a.s., from 270 to 195 thousand. Even then, it's a few tens of percent for free.
A different model for each job
Not every job needs the smartest — and most expensive — model. Reading numbers off a table is something a cheap model handles as reliably as an expensive one. Writing an investment thesis is another league.
So that's how I do it. Sonnet runs extraction, Opus the thesis, Haiku the small stuff. Price per million tokens (input / output):
- Haiku: $1 / $5
- Sonnet: $3 / $15
- Opus: $5 / $25
Running Opus on work Haiku can do is five times the price for the same result. On top of that, I control the models from one place — today I run on Claude, but I'm not tied to anyone. Something better or cheaper shows up, I swap it.
I read a public filing once — for everyone
Data from the commercial register is public and immutable. The 2023 filing won't change anymore.
So I read it just once. I store the result, and when someone else opens the same company, they get it done — for free. The more Misano is used, the more filings are already done, and the lower the cost of the next analysis.
This is strictly public register documents. Your own data is never shared with anyone.
I meter every job
I don't watch cost by eye. Every company has a $2 cap on processing, plus daily limits. By the way, when a data transfer fails, the job costs the user $0. That keeps our incentive fully aligned — flawless parsing.
On top of that, more small details that save tokens. I cache the shared part of the prompt across years and companies in a single batch, so I'm not paying again and again for the same thing. And I take the most recent year's data from ChytryRejstrik.cz for a flat fee instead of parsing it — which starts to pay off across thousands of companies a month.
What it adds up to
Totalled on Barrandov Studio a.s. In your ChatGPT or Claude, that filing as images on a top model would run about $0.75 — and again every time. In Misano, thanks to text instead of images and a cheaper model for extraction, $0.11. And if someone already opened it before me, $0.
$0.75 → $0.11 on the first read, repeats free. The average per company lands around −90% — exactly where I was aiming. And that's a single filing. A fund runs through hundreds in a year.
What you get out of it
Two things. A price I don't have to bury a fat AI bill in — Misano costs a fraction of what building the same analysis by hand would. And a product that will hopefully still be here in a few years — because tools that just wrap a single LLM with no added value, you'll probably end up replacing with your own Claude or ChatGPT.
Try it
You'll see it best yourself. Click Get started, enter a real company, and watch the Price column on every job — actual numbers, not estimates. Then compare it with what you'd leave behind in your own ChatGPT for the same work.
How I calculate this
The numbers above are from real usage, not estimates. A token = input plus output actually consumed on a job, prices per Anthropic's list as of 15 June 2026. The 'in your ChatGPT / Claude / Gemini' comparison = the same filing uploaded as PDF pages to a top model, with no cache and no sharing between users.
And the −90%, to be fair about it: the first read of one filing falls from ~$0.75 to ~$0.11, which is −85%. But because I store every public filing and the next user gets it for free, the average cost per company across repeats drops toward −90%. Your numbers will vary by company, number of years, and how clean a text layer the filing has.
Thanks for reading this far — and if something's missing in Misano, write to me at hello@misano.ai.