onrup

Tool

Fine-tuning cost calculator

How much does a fine-tuning run cost?

Cost is the hourly rate of the GPU class multiplied by wall time, and wall time is driven by total training steps — dataset size divided by effective batch size, times epochs. Pick a model, a dataset size and a class below for an order-of-magnitude estimate, then use the forecast endpoint for a real number.

Estimated run

Serving, always warm

per 30 days, one replica

Fit

An estimate, not a quote. It assumes an effective batch of eight and scales throughput from the class and the model’s active parameter count. Real wall time depends on your sequence length distribution and adapter settings, which is what the forecast endpoint accounts for.

Where estimates usually go wrong

The full arithmetic, and where budgets go wrong →

Frequently asked questions

How accurate is this estimate?

It is an order-of-magnitude guide, not a quote. Real wall time depends on sequence length distribution, batch size, adapter configuration and memory-saving settings, none of which this page knows. The forecast endpoint gives a real number against a real dataset before you commit.

Why does a faster GPU sometimes cost less overall?

Total cost is the hourly rate multiplied by wall time. A class at twice the rate that finishes in a third of the time is cheaper. The effect is strongest when the cheaper option requires four-bit training, which is meaningfully slower per step.

Is training the main cost of a model?

Usually not. Training is a handful of bounded runs; serving is every request forever. Use the serving figure below as the number that actually matters for a model that reaches production.

Last verified 6 August 2026.

Start with the free tier

A magic link creates your account, your tenant and your first API key. No card until you ask for compute.