Fine-Tuning Cost Estimator

Data preparation, training runs and monthly serving costed together, and compared against prompting the base model with a longer prompt.

Data preparation

Training examples

Tokens per examplePrompt and completion together.

Human hours to build the setCollecting, cleaning, labelling, reviewing.

hours

Loaded hourly costWhat an hour of that person's time actually costs you.

$ / hour

Training

Epochs

Training rateUS dollars per 1,000,000 tokens of training data processed.

$ / 1M

Runs before you ship oneNobody fine-tunes once. Hyperparameters, data fixes, a rebuild after the eval fails.

Serving the fine-tune

Requests per month

Input tokens per requestWith the fine-tune, the prompt gets shorter. This is the short version.

Output tokens per request

Fine-tuned input rateUS dollars per 1,000,000 tokens

$ / 1M

Fine-tuned output rateUS dollars per 1,000,000 tokens

$ / 1M

Dedicated hosting rate0 if your fine-tune runs on shared capacity and is billed per token only.

$ / hour

Hosted hours per month730 for a month of always-on dedicated capacity.

hours

The alternative — prompting the base model instead

Extra prompt tokens without the fine-tuneThe instructions and few-shot examples the fine-tune replaces. This is where the saving comes from.

Base model input rateUS dollars per 1,000,000 tokens

$ / 1M

Base model output rateUS dollars per 1,000,000 tokens

$ / 1M

One-time cost before a single production request

$3,416

Serving the fine-tune saves $1,350 a month against prompting the base model, so it pays back in 2.5 months. Over twelve months: $16,376 fine-tuned against $29,160 not.

Training tokens per run: 9,000,000
Cost per training run: $72.00
Training, 3 runs: $216.00
Data preparation, human time: $3,200
One-time total: $3,416
Prep as a share of one-time: 93.7%
Serving, per request (fine-tuned): $0.0036
Dedicated hosting, per month: $0
Serving, per month (fine-tuned): $1,080
Serving, per request (base + long prompt): $0.0081
Serving, per month (base + long prompt): $2,430
Monthly difference: $1,350
Payback: 2.5 months
First year, fine-tuned: $16,376
First year, base model: $29,160

What this assumes: training cost is billed per million tokens of data processed, and tokens processed is examples × tokens per example × epochs. Providers that bill fine-tuning by GPU-hour instead will not match this — use the GPU cost tools for that shape. The comparison assumes the fine-tune buys you a shorter prompt and nothing else: same output length, same number of requests, same quality. If it also removes a validation call or a retry, that saving is real and is not counted here. Human time is priced at the loaded hourly figure you entered, which is the number people leave out and which is frequently the largest single line above. Evaluation runs, storage of the training set, and the engineering time to wire the whole thing up are excluded.

The reason fine-tuning budgets are wrong is not the training bill. The training bill is usually the smallest of the three numbers on this page, and often small enough to disappear. The costs that matter are the human hours spent building a dataset good enough to be worth training on, and the ongoing serving cost of the result — which, if the fine-tuned model carries a rate premium over its base, can quietly exceed everything you saved on prompt length.

That is why this calculator insists on the comparison. A fine-tune is not free money; it is a capital expense you are making in order to reduce a per-request one, and the mechanism is almost always prompt length. You are replacing two thousand tokens of instructions and examples, sent on every single request forever, with weights. Whether that trade works is a division: one-time cost divided by monthly saving. Under a few months and it is obviously right. Over a year and your requirements will have changed before it pays back, at which point you are re-running data prep too.

Three runs is the default here because one is a fantasy. The first run reveals that the dataset has a formatting inconsistency, the second overfits, the third is the one you ship. If you are also evaluating against a held-out set, add those inference costs. And if the honest answer to "why are we doing this" is latency or a behaviour that prompting cannot reach rather than cost, this page will tell you the fine-tune never pays back — which is fine. Just do not sell it internally as a cost saving when the arithmetic says it is not one.

Fine-Tuning Cost Estimator

The reasoning behind this

Related tools