Everyone in this market has an incentive to publish a number that flatters them. Ours is to publish only the numbers we can point at, and to say nothing at all when we cannot.
We open the provider's own pricing page and read the numbers off it. Not a model catalogue, not a comparison site, not a blog post summarising the launch — the page the vendor maintains, because that is the one they will bill you against.
Four things get recorded with each number: the channel it was quoted on, the processing tier (standard, batch, fast, off-peak), the date we read it, and the URL. Drop any one of those and two prices for the same model stop being comparable — which is how budget spreadsheets end up wrong by a factor of seven.
Prices live in a snapshot store. Each pass is diffed against the last, so a change shows up as a dated entry in the price watch instead of a quiet edit. If you bookmarked a number last month, you can find out what happened to it.
When a fetch fails, nothing is overwritten. The old value keeps its old date. We would rather show you a stale number with an honest timestamp than a fresh guess.
Some outbound links — mostly GPU rental — are affiliate links and earn a commission if you sign up. They are labelled where they appear, and full details are on the disclosure page.
What that never touches is the price table. The rates printed here are the providers' published rates regardless of whether a link earns anything. A directory that quietly reorders by commission is not a directory, it is an ad.
If a number here disagrees with the provider's page, email pipeitouzi@gmail.com. We re-read the source, correct it if we are wrong, and date the correction. Corrections are logged rather than quietly edited — the history is the useful part.
Standing caveat: providers change pricing without notice and sometimes without announcing it. Confirm on the provider's own page before you commit budget to anything you read here.