Fine-tuning cost
Two bills, one project: a one-time training run and a monthly serving tab. The trap is assuming the training bill is the big one — for most teams, serving dwarfs it within a quarter.
The formula
Training-rate default is an editable placeholder in the typical hosted range ($2–4/1M tokens as of ) — vendors change it, so paste yours. Serving bills at base-model per-token rates via OpenRouter's daily pull. Fine-tuning pays off only when it lets you shrink prompts (bake instructions in) or downgrade the base model — model the saving on the chatbot calculator for comparison.
Two bills on two different clocks
Fine-tuning is the only calculator on this site where the headline number can be genuinely misleading, because it is really two numbers that do not belong in the same sentence.
| Bill | Paid | Formula |
|---|---|---|
| Training run | Once | examples × tokens/example × epochs × training rate |
| Serving | Every month, forever | input tokens × in-rate + output tokens × out-rate |
Dividing the training cost by twelve and calling it a monthly cost is arithmetically possible and analytically useless. It hides the fact that the serving bill keeps arriving long after the training invoice is paid, and it makes a project that is getting more expensive each month look flat.
Why serving quietly becomes the larger number
The first twelve months are the honest unit of comparison. A tuning run in the low hundreds of dollars, serving a modest production workload, passes it within a few months and never looks back. The break-even between training and serving is not a break-even at all — it is a crossover you should calculate before you start, so nobody is surprised in month four.
Which raises the only question that matters: what does the tuned model give you that the base model did not, and can you put a monthly number on it?
Fine-tuning pays only when it shrinks something
There are exactly two ways tuning reduces a bill, and both are about shrinking:
- Prompt shrinkage. Behaviour, formatting rules and long few-shot examples get baked into the weights, so every production call sends fewer input tokens. The saving is the reduction in prompt size multiplied by your call volume.
- Model downgrade. A smaller tuned model matches the accuracy of a larger general one on your narrow task. The saving is the rate difference between the two models across all your traffic.
Both are calculable, and both should be computed on the chatbot calculator or the relevant workload calculator before the training run is booked. If neither effect applies — the prompt stays the same length and the base model tier stays the same — then fine-tuning is a pure addition to your bill, whatever quality it buys you.
Where teams get this wrong
- Justifying the project on the training invoice. It is the smaller half and the only half that stops.
- Omitting annotation labour. Producing clean examples is human work, usually several times the training cost, and it never appears on a vendor invoice.
- Forgetting that tuning invalidates your evaluation baseline. A tuned model needs its own evals, its own regression suite, and it needs to be re-checked every time the base model is upgraded.
- Tuning to teach facts. Facts belong in retrieval, where they can be updated in minutes. Facts baked into weights go stale and require the whole training cycle again.
- Assuming the vendor's rate is stable. Training prices move; the default here is an editable placeholder for exactly that reason.
The cost that comes back: re-tuning
Training is described as a one-time cost, and it is — per training run. It is not one-time per project. Every time the base model you tuned is upgraded or deprecated, the tuned weights do not carry over: you retrain, re-evaluate and redeploy. Providers move fast enough that over a product's life this happens more than once.
There is a second recurring cost that is even easier to miss. A tuned model invalidates your evaluation baseline: your prompts, your test set and your quality thresholds were all calibrated against a specific model version, and retraining means re-running that calibration — because a new model can score the same average and fail on entirely different cases.
Budget for both. A tuning project costed on a single training run, with no re-tuning line, is a project carrying an unbounded future cost.