How to estimate what a fine-tune costs

No table can tell you how long your job runs. It depends on your model, your data, your sequence length and your settings. So measure it. Start the GPU you plan to rent, run a few hundred training steps, and note how long they took.

Divide the steps in a full run by the steps you timed, and multiply by the time. That is the length of one run in hours. This site prints no training speeds, because a speed measured on another job is not yours.

Multiply the hours by the number of GPUs to get GPU-hours. Multiply the GPU-hours by the price per GPU-hour. On 12 Nov 2026, the median listed H200 price here was $5.40 per GPU-hour, so a run of ten GPU-hours costs ten times that figure. On the RTX Pro 6000 the median was $2.20, and the same ten GPU-hours cost ten times that instead.

Use a price you can start, not the lowest one printed. The tables here mark each listing with whether it can be started today.

Nobody pays for one clean run. You pay for a run that crashes, a run with the wrong learning rate, and a few runs to compare settings. Count them honestly. If you expect one final run and four tries before it, budget for five, and shorten the tries where you can.

Storage for the dataset and the checkpoints is billed by the gigabyte for as long as it exists. Data transfer may be billed when you move the finished model out. Idle time is billed whenever the machine is on and not training. Write each as its own line in the estimate.

Runcrate bills by the second on most GPUs and never bills booting, so the timing run and the short tries cost what they used. Runcrate can start an H200 now from $4.67 per GPU-hour. One prepaid balance pays for the GPU and the storage.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.