Fine-Tuning Cost Calculator

Fine-tuning cost splits into two entirely different pricing models depending on where the training runs, and this calculator switches between them based on which model you pick. For the API-based options — GPT-4o mini, GPT-4o, and GPT-3.5 Turbo — the tool separates your dataset into a training portion and a validation portion by the split percentage, but both portions still get priced at the same per-million-token training rate and summed into one bill: the split changes how the total is reported (a "validation examples" count shows up as its own line), but it does not change the total number of tokens billed, since every example — training or validation — passes through training at least once per epoch. In other words, raising the validation split reshuffles which examples are labeled "training" versus "validation" without lowering API cost; GPT-4o's $25-per-million rate versus GPT-4o mini's $3 is what actually swings the total. The self-hosted options (Llama 3, Mistral) work differently: GPU-hours are estimated only from the training portion — about one A100 GPU-hour per 10,000 training examples per epoch for Llama 3, slightly less for Mistral — at an assumed $1.50/hour spot price, so a larger validation split does reduce estimated self-hosted compute cost, unlike the API path.

On top of raw training cost, the calculator adds a dataset preparation estimate — roughly 2 minutes of labeling time per example at $50/hour — which for smaller datasets often exceeds the training cost itself. A fixed 40% prompt-token savings figure is also shown, reflecting the common (but not guaranteed) benefit that a fine-tuned model needs a shorter prompt at inference. Confirm current rates against your provider's live pricing page before committing budget.