Fine-Tuning Cost Calculator

Estimate LLM fine-tuning costs from dataset size, epochs, and model choice for OpenAI API and self-hosted options.

Training examples

Avg tokens per example

Training epochs

Model

Validation split (%)

Tab to move between fields

Total estimated cost ($)

$1,671.17

≈ 13 pairs of sneakers

Training/compute cost ($)$4.50

Dataset prep cost ($)$1,666.67

Training tokens (millions)1.35

Est. training time (hours)1.4

Dataset prep hours33.3

Validation examples100

Prompt token savings (%)40%

Inference Price Per1M$0.30

How to Use This Calculator

Enter the number of Training Examples (prompt-completion pairs) in your dataset — typically 500–5,000 for most fine-tuning tasks.
Set average Tokens per Example (prompt + completion combined) and the number of Training Epochs (2–4 is typical).
Select the Model — API-based fine-tuning (GPT-4o mini, GPT-3.5) or self-hosted (Llama, Mistral).
Set the Validation Split percentage to hold out data for evaluation.
Review Total Estimated Cost, Training Tokens, Estimated Training Time, and Dataset Prep Hours to plan your fine-tuning project.

How the result changes with Training examples

Training examples: Number of training examples (prompt-completion pairs) in your dataset.
Avg tokens per example: Average total tokens (prompt + completion) per training example.
Training epochs: Number of passes through the training data. OpenAI recommends 2-4 for most tasks.
Model: Select the model
Validation split (%): Percentage of examples reserved for validation. Typical: 10-20%.

Total estimated cost ($): Training cost + dataset preparation cost.
Training/compute cost ($): API fine-tuning fee or self-hosted GPU compute cost.
Dataset prep cost ($): Estimated data labeling cost at $50/hr (~2 min per example).
Training tokens (millions): Total tokens processed during training across all epochs.
Est. training time (hours): Estimated wall-clock training time.
Dataset prep hours: Estimated hours for manual data preparation and labeling.
Validation examples: Number of examples held out for validation.
Prompt token savings (%): Estimated inference prompt savings vs. few-shot prompting (~40%).

Identify Input Parameters
4 parameters
Training examples = 1000, Avg tokens per example = 500, Training epochs = 3, Model = 0 = 5 input(s) provided
Calculate Total estimated cost
Total estimated cost = trainingCost + datasetPrepCost
1671.17 = $1,671.17
Calculate Training/compute cost
Training/compute cost
4.5 = $4.5
Calculate Dataset prep cost
Dataset prep cost = datasetPrepHours * 50
1666.67 = $1,666.67