Fine-Tuning Cost Calculator
Estimate the training cost and ongoing inference cost of fine-tuning an OpenAI language model.
Did this tool work for you?
How to use this calculator
- 1
Select the base model you plan to fine-tune.
- 2
Enter the number of training examples and their average token length.
- 3
Enter the number of epochs (training passes). More epochs = more cost and sometimes better results.
- 4
Enter your expected monthly inference volume to see the ongoing cost alongside the one-time training cost.
Frequently asked questions
What formula gives the training cost?
Training tokens = examples × average tokens per example × epochs. Training cost = training tokens ÷ 1,000 × the base model’s training rate. Tokens are billed on every epoch, so doubling the epochs doubles the training cost.
What happens if I enter fewer than one epoch?
The code treats any value below 1, including zero, as 1 epoch. A cleared epochs box becomes 0 on the page, so it also counts as 1. Fractions above 1 are used as entered.
How is monthly inference cost estimated?
The form asks for one token figure covering input and output, so the tool splits it evenly: half priced at the input rate and half at the output rate. Monthly inference = requests × (half the tokens × input rate + half × output rate) ÷ 1,000. Skewed workloads, such as long prompts with short answers, will differ from the estimate.
What are the rates in the tool?
Per 1,000 tokens, training is 0.003 for GPT-4o-mini, 0.008 for GPT-3.5 Turbo and 0.025 for GPT-4o, equal to $3, $8 and $25 per million. The inference rates for each model are in the table on this page. They are figures built into this tool, so check the provider’s fine-tuning pricing page before budgeting.
Which base models can I choose?
Three entries: GPT-4o-mini, GPT-3.5 Turbo and GPT-4o. The tool does not know which base models a provider currently allows you to fine-tune, so confirm availability separately.
What does the six-month total contain?
Training cost once, plus six times the monthly inference cost. It does not include data preparation, evaluation runs, validation data, failed or repeated jobs, or any hosting or storage fee for the tuned model.
Does the calculator tell me whether fine-tuning beats prompting?
No. It prices the fine-tuned path only. To compare, estimate the same traffic with a base model in a prompt cost calculator and set the difference against the training bill. It also cannot predict quality, which is not a function of the number of examples alone.
Why is the training rate shown with three decimals?
Rates are small dollar amounts per 1,000 tokens, so three decimals keep 0.003, 0.008 and 0.025 distinct. Money results are rounded to two decimals.
Fine-tuning cost calculator: training spend, monthly inference and six-month total
Training cost formula
Training tokens are training examples × average tokens per example × epochs. Training cost is those tokens ÷ 1,000 × the base model’s training rate. Epochs below 1, including 0 (which is what a cleared box becomes), are treated as 1, and tokens are billed on every pass, so doubling the epochs doubles the training line. The cost is one-time and is incurred again whenever you retrain on a new or larger dataset. Average tokens per example should include both the prompt and the target answer of each training record.
Rates stored in the code
These figures are built into this tool and are expressed in US dollars per 1,000 tokens. The list has three base models and does not track which models a provider currently lets you fine-tune. Check the provider’s fine-tuning pricing page before budgeting.
| Base model | Training | Inference input | Inference output |
|---|---|---|---|
| GPT-4o-mini | $0.003 | $0.0003 | $0.0012 |
| GPT-3.5 Turbo | $0.008 | $0.003 | $0.006 |
| GPT-4o | $0.025 | $0.0025 | $0.01 |
Worked examples
Fine-tuning scenarios. The columns show Scenario, Training tokens, Training cost, Monthly inference and Six-month total, one line per case.
| Scenario | Training tokens | Training cost | Monthly inference | Six-month total |
|---|---|---|---|---|
| GPT-4o-mini, 1,000 examples × 500 tokens × 3 epochs, 10,000 requests × 300 tokens (defaults) | 1,500,000 | $4.50 | $2.25 | $18.00 |
| GPT-4o, 2,000 × 800 × 3 epochs, 50,000 requests × 600 tokens | 4,800,000 | $120.00 | $187.50 | $1,245.00 |
| GPT-3.5 Turbo, 500 × 400 × 4 epochs, 5,000 requests × 200 tokens | 800,000 | $6.40 | $4.50 | $33.40 |
How inference is estimated
Because the form asks for a single token figure covering input and output, the code splits it 50/50: half the tokens at the input rate, half at the output rate. In the GPT-4o run, 600 tokens per request becomes 300 in and 300 out. A workload with a long prompt and a short reply will cost less than shown on models where output is priced higher, and long generations will cost more. The GPT-4o run also shows the typical shape of the result: a six-month total of $1,245.00 against a $120.00 training bill, so usage, not training, drives the budget.
What the totals leave out
Data preparation, evaluation runs, validation-file tokens, failed or repeated jobs and any hosting or storage charge for a tuned model are not in the estimate. The tool also does not price the untuned alternative. To judge whether tuning pays off, run the same traffic through a base-model calculator, compare it with the inference line here, and weigh the training bill against the difference over the period you care about.
Learn more from an authoritative source:
OpenAI Platform DocsAI Image Generation Cost Calculator
Calculate monthly costs for AI image generation with DALL-E 3, Stable Diffusion, or Midjourney.
RAG Pipeline Cost Calculator
Estimate the monthly cost of a Retrieval-Augmented Generation (RAG) pipeline including indexing, embeddings, and LLM inference.
GPU Cloud Cost Calculator
Compare GPU rental costs across AWS, GCP, Azure, Lambda Labs, and Vast.ai for ML training and inference workloads.
AI Chatbot ROI Calculator
Calculate the return on investment of deploying an AI chatbot for customer support, including payback period and annual ROI.
LLM Latency Calculator
Estimate expected LLM response latency — time-to-first-token and total response time — for planning real-time AI applications.
AI Token Counter
Estimate the number of tokens in your text for GPT-4, Claude, Gemini, and other LLMs. Useful for staying within context limits.
Results are estimates for informational purposes only and do not constitute professional financial, medical, legal, or technical advice. Read full disclaimer →