Pricing
How GPU-based training is billed against your credit balance.
Training runs on real GPUs provisioned on-demand. Cost is charged from your credit balance.
How training is billed
The charge is based on GPU-hours at the standard per-GPU rate for the hardware your job runs on.
You are charged the greater of the token-based estimate and the actual GPU time used, so a run that finishes fast still covers its estimate, and a long run is billed for the real time it consumed.
Check the estimate before you run
The New Fine-Tuning Job screen shows a live cost estimate as you configure a job — always check it against your balance before clicking Start.
| Component | Detail |
|---|---|
| Rate | GPU-hours × per-GPU hourly rate |
| Final charge | The greater of the token-based estimate and the actual GPU time used |
| Paid from | credit balance |
- LoRA and QLoRA jobs generally cost less — they train fewer parameters and finish faster on smaller GPUs.
- Full fine-tunes cost more — every weight updates, which needs bigger GPUs and longer training time.
- Cancelling a job early still bills for the GPU time already consumed, not the original estimate.
Note
GPU rates vary by hardware. See GPU instance pricing for the per-GPU rates.