AI Fine-Tuning Cost Calculator
Work out what it costs to fine-tune a language model through an API or on your own GPU, what it costs to run each month, and how soon it pays for itself against a longer prompt on the base model.
500,000 training tokens in total.
Example price only — check your provider's pricing page for the real rate.
What one GPU costs to rent (or run) an hour.
For all the GPUs together. Measure it on a short test run.
Rules of thumb for English text. Code and other languages often use more tokens.
Extra examples held back to check the model each epoch.
Some providers bill at least this much.
Count retries and retraining on new data.
Running the fine-tuned model
Only if your provider charges to keep the model deployed.
730 hours is a whole month.
With the fine-tuned model.
Compared with prompting the base model
Instructions and examples the base model needs that the fine-tune has learned.
Example prices only — enter what your provider charges.
How your fine-tuning cost is worked out
Count the tokens, times the epochs, times your price. Then compare running costs.
- 1
Count the tokens
1,000 × 500= 500,000 tokens
- 2
Times the epochs
500,000 × 3= 1.5M billed
- 3
Training cost
1.5M × $8.00 ÷ 1M= $12.00 once
- 4
Running and payback
Usage: in AdvancedYear 1 = $12.00
Read more: the formula, a worked example and notes
Fine-tuning teaches an existing model from your own examples. Through an API you pay for every token the model trains on, and it reads your whole dataset once per epoch. So the training cost is a count: dataset tokens × epochs × your price per million training tokens. After that you pay to use the model each month, and sometimes an hourly fee to keep it deployed. It pays off when the fine-tuned model lets you send much shorter prompts than the base model needs.
The formula
tokens per example = words ÷ words per token (or characters ÷ characters per token), rounded
dataset tokens = examples × tokens per example (or the whole dataset size in tokens)
validation tokens = dataset tokens × validation %
billed training tokens = (dataset tokens + validation tokens if billed) × epochs
training cost = billed tokens ÷ 1,000,000 × price per 1M (at least the minimum charge)
on your own GPU: hours = tokens ÷ tokens a second ÷ 3,600, and cost = hours × GPUs × price per GPU-hour
running cost a month = hosting fee × hours + requests × (input tokens × input price + output tokens × output price) ÷ 1,000,000
base model a month = requests × ((input + extra prompt tokens) × base input price + output tokens × base output price) ÷ 1,000,000
saving a month = base model a month − running cost a month
break-even = training cost ÷ saving a month
first year = training cost × training runs + 12 × running cost a month
Worked example
Take 1,000 examples of 500 tokens: 1,000 × 500 = 500,000 tokens. Trained for 3 epochs, that is 500,000 × 3 = 1,500,000 billed tokens (the 10% validation set, 50,000 tokens, isn't billed here). At a price of 8 per 1M, training costs 1.5 × 8 = 12.00 once, or 4.00 per epoch. If validation tokens were billed, it would be 550,000 × 3 = 1,650,000 tokens and 13.20.
Running it (Advanced): 100,000 requests a month with 200 input and 100 output tokens. On the fine-tuned model at 0.3 and 1.2 per 1M, one request costs (200 × 0.3 + 100 × 1.2) ÷ 1,000,000 = 0.00018, so the month costs 100,000 × 0.00018 = 18.00 (no hosting fee). The base model needs 1,500 more prompt tokens: (1,700 × 0.15 + 100 × 0.6) ÷ 1,000,000 = 0.000315 a request, or 31.50 a month.
The saving is 31.50 − 18.00 = 13.50 a month, so the 12.00 training cost pays back in 12 ÷ 13.5 = 0.889 months, about 27 days. The first year costs 12 + 12 × 18 = 228 against 12 × 31.50 = 378 for the base model. To pay back within a year you need 1 ÷ 0.000135 = 7,407.4, so 7,408 requests a month. On your own GPU at 1,000 tokens a second, the same 1,500,000 tokens take 1,500 seconds (0.4167 hours), which at 2 an hour is about 0.83.
Size rules used
| You enter | Tokens | Rule (editable in Advanced) |
|---|---|---|
| 300 words | 400 | 0.75 words per token |
| 2,000 characters | 500 | 4 characters per token |
- An hourly hosting fee changes everything. A fee of 1.50 an hour for 730 hours is 1,095 a month, far more than the token saving in the example. If your provider charges one, the fine-tune only pays when you have a lot of traffic.
- Fine-tuned models can cost more per token than their base model. Put both prices in to see the real saving.
- Plan for more than one run. First tries often need a fix to the data or the epochs. Put the runs you expect in "Training runs in the first year".
- Try prompting first. A good prompt, or prompt caching for a long system prompt, can be cheaper than fine-tuning for low volumes.
No provider prices are built in: enter the rates you pay. The example price buttons only show a range and aren't any provider's current price. Rules checked 2026-10-17.
Common questions
How much does it cost to fine-tune a model?
Training tokens × epochs × your price per 1M training tokens. 1,000 examples of 500 tokens is 500,000 tokens; 3 epochs makes 1,500,000 billed tokens; at a price of 8 per 1M that is 1.5 × 8 = 12. Check your provider's pricing page for the current rate, and whether it adds a minimum charge or a fee for hosting the model.
What is an epoch?
One full pass of the model over your training data. With 3 epochs the model reads every example three times, so you pay for three times the tokens. More epochs can help a small dataset, but too many can make the model copy your examples instead of learning from them.
Are validation tokens billed?
It depends on the provider. Some charge for the validation set as it is checked each epoch, some don't. Use the "Validation tokens billed" switch in Advanced to see both.
When does fine-tuning pay for itself?
When it lets you send shorter prompts. A fine-tuned model has learned your instructions and examples, so each request can drop them. If you save 1,500 prompt tokens on 100,000 requests a month, that saving soon pays back a one-off training cost, unless a fine-tuned model costs more per token or carries an hourly hosting fee.
How long does it take to fine-tune on my own GPU?
About tokens processed ÷ tokens a second. 1,500,000 tokens at 1,000 tokens a second is 1,500 seconds, about 25 minutes, or 0.4167 GPU-hours. At 2 an hour on one GPU that is about 0.83. Your speed depends on the model size, the GPU and whether you use LoRA.
How many tokens are in my examples?
For English, about 0.75 words per token and about 4 characters per token, so 300 words is about 400 tokens. Chat-format examples add a few tokens per message for roles and separators. Count a sample with a token counter for a closer figure.
Related calculators
Made by Zulu Calculator, part of AlphaZulu LLC. Formulas are shown in full and rule figures are checked against official sources; see how we build calculators.