Fine-tune for behaviour, not for knowledge
Fine-tuning teaches a model how to respond — format, tone, structure, classification boundaries. It is a poor and expensive way to teach it what is true. For facts that change, use retrieval (RAG). Teams that fine-tune to inject knowledge usually end up retraining every time the data changes, which is the most expensive possible loop.
The hidden cost: hosted inference premium
Providers commonly charge more per token to serve a fine-tuned model than the base model — often around 2×. If your prompt savings are small, that premium quietly eats the benefit and you end up paying more forever in exchange for a one-off training bill.
Budget for failed runs
First fine-tunes are almost never the one you ship. Data cleaning issues, wrong epoch counts, and overfitting mean 2–4 runs is realistic. Multiply your single-run estimate accordingly, and add engineering time — which usually dwarfs the compute cost at small scale.
Related guides
Long-form playbooks on the same topic, written by the RevenueLab editorial team.
FAQ
How much does it cost to fine-tune an LLM?
A typical style/tone tune — 5,000 examples at 600 tokens, 3 epochs, at $8 per million training tokens — costs about $72 per run, so roughly $215 across three runs. Large domain tunes on bigger models run into the thousands.
Is fine-tuning cheaper than prompting?
Only at volume. Fine-tuning saves prompt tokens on every call but typically adds a per-token inference premium. Below roughly 50,000–100,000 calls per month, a well-engineered prompt plus caching usually wins.
Fine-tuning vs RAG — which do I need?
RAG for knowledge that changes; fine-tuning for behaviour that doesn't. Most production systems use both: a fine-tune for output format and tone, retrieval for the facts.
How many examples do I need?
500–1,000 clean examples is enough for tone and format. Classification typically needs 50–100 examples per class. Data quality beats quantity — 500 excellent examples outperform 5,000 noisy ones.
What about LoRA / open-source fine-tuning?
LoRA on rented GPUs shifts the cost from per-token to per-GPU-hour. A small LoRA often trains in 1–4 GPU-hours; price it in the GPU cloud cost calculator, then compare against the hosted number here.
How many epochs should I use?
Three is the common default. More epochs increase cost linearly and raise overfitting risk — the model starts parroting training examples verbatim instead of generalising.
How this calculator is built
Independently maintained
Written by Sam Doshi and the RevenueLab editorial team. We don't sell the data feeds this tool is built on.
Sourced from primary data
Benchmarks come from public AdSense / Stripe / IRS disclosures and reader-submitted data — never third-party "$X per view" claims. Full methodology.
Last editorial review
Reviewed on a rolling quarterly cycle. Dated reviews are published on the methodology record for each calculator.
Editorial standards
See our editorial policy and disclaimer. Results are estimates, not advice.