Home / Tools / Cost calculator

What does this fine-tune cost to run?

Training cost is tokens, not examples. Enter what you have; the estimate does the rest.

Run cost idle
Model
GPU
Method
Your price per GPU-hour ($)
Training examples
Tokens per example
Epochs
Sequence length
$ — per run
Training tokens—
Throughput—
GPU hours—
VRAM needed—
Fits this GPU—

Throughput is a 6·N flops/token estimate with a measured efficiency factor. Real runs vary by 20-40% with framework, sequence packing and dataset shape.

Why tokens, not examples

A training step costs roughly 6 x N floating point operations per token, where N is the active parameter count. So the bill scales with tokens seen, not with how many rows your JSONL has.

Five thousand examples at 200 tokens is a quarter of the compute of five thousand examples at 800 tokens — but people compare the two as if they were the same job.

Where the estimate drifts

  • Sequence packing can raise effective throughput 30-60% by removing padding.
  • Framework choice matters more than the GPU. Unsloth's kernels and a stock TRL loop differ by more than a 4090 and an A100 on small models.
  • Data loading becomes the bottleneck when the dataset does not fit in page cache.
  • Checkpoint saving can add minutes per epoch on network storage.

The number that actually matters

Rent by the hour and the training run is almost always cheap: a few dollars. The expensive parts are the labelling, the failed attempts and the evaluation. Budget accordingly — see the full guide.