How this works
About
A reference for people building with AI who are not machine-learning engineers: what the models cost, which one suits a given job, and what the vocabulary means.
Why it exists
Model pricing is published, and it is still genuinely hard to answer two ordinary questions: what will this cost me, and which one should I use. Vendor pages are marketing surfaces, comparison sites copy each other without citing anyone, and almost nobody models what the prompt cachecosts to write — which on agentic work is frequently the largest line on the bill.
So the useful thing is not another table. It is a table where every figure says where it came from and when, and where the gaps are visible instead of filled in.
How it is paid for
It is not. There are no affiliate links, no sponsored placements, no commission on any choice, and nothing on this site earns money if you pick one model over another. If that changes it will be stated here first, in these words, before anything else changes.
What it will not do
- Publish a price that is not on the vendor’s own page, verbatim.
- Turn a figure a vendor declines to publish into a zero or a guess.
- Pick a winner when two sources disagree.
- Hide a source that refused us.
- Rank models as “best” without saying best at what.
How it is built
A Python pipeline with no dependencies reads vendor pricing pages and two public price databases once a week, applies a fixed merge policy, and writes the data this site is built from. It runs in a throwaway build container, never in your browser, so reading this page does not send a request to any vendor.
The site itself is static. There is no tracking, no analytics, no cookies and no third-party scripts — a strict content policy forbids them, which is also why there are no web fonts loaded from anywhere else.
Mistakes
A wrong price is the worst thing that can be here, and the fastest way to fix one is to be told about it. Say which model, what the vendor’s page shows, and link it.