About
A bench for fine-tuning arithmetic
This site turns a model size and a token count into GPU-hours and dollars, and shows every step so you can disagree with any of them.
What it does
It applies a standard rule of thumb for training compute, divides by the throughput a GPU maker prints, scales by an efficiency you can edit, and multiplies by the price providers list per GPU-hour. The scenario pages run the same arithmetic for sixteen common cases.
What it will not claim
It does not claim a real job will take this long or cost this much. We did not run or read any benchmark for these pages. GPUs whose maker prints only a sparse figure are left out instead of being halved, and the coverage page says which and why.
Who the numbers come from
Throughput figures are quoted from the makers’ own pages with a link. Prices are published on-demand list prices read from each provider’s page, with the date shown.
Who runs this
Fine-Tune Cost Lab is published by the team behind Runcrate, an AI cloud. We list Runcrate in the tables where it has a price, marked as ours, and sort by price like every other row.