The training run is often the smallest line item. Budget for data work, evaluation, repeated experiments, hosting the result and retraining when the base model changes. Get current prices from your provider or cloud before committing, because they change quickly.
Cost buckets
- Data. Collecting, cleaning, labelling and reviewing examples. This is mostly people time and is often the largest cost.
- Training compute. GPU time or provider training fees.
- Experiments. You will run several jobs, not one: data variants, settings, ranks.
- Evaluation. Building the eval set and running it after each change, including model graded evals that cost tokens.
- Serving. Hosting your own model or paying a provider to serve a tuned one.
- Upkeep. Retraining when data shifts, the base model is retired, or safety testing is repeated.
What the research says about compute scale
Papers give a sense of scale, with their own hardware and dates. The phi-1 report describes a 1.3 billion parameter model trained in 4 days on 8 A100 GPUs. QLoRA's authors say training their Guanaco models took about 24 hours on a single GPU, and that the approach lets a 65B model be tuned on one 48 GB GPU. LoRA's authors report it can cut GPU memory needs by 3 times compared with fully fine-tuning GPT-3 175B with Adam.
Those are research setups from different years. We are not quoting prices, because rates vary by provider and region and change often. Multiply your own measured GPU hours by the current rate.
Ways to keep the bill down
- Start with an adapter method such as LoRA or QLoRA.
- Run small pilot runs on 100 examples first, then double.
- Use a small base model where it is good enough, see small language models.
- Cap the number of experiments and write down what each one tests.
- Cache eval outputs for unchanged prompts.
A simple budgeting template
| Item | Quantity | Unit cost | Notes |
|---|---|---|---|
| Labelling hours | your estimate | your rate | Include review |
| Training hours | pilot x runs | current rate | Check date |
| Eval tokens | per run x runs | current rate | Include judge model |
| Serving | requests per month | current rate | Compare with a hosted baseline |
| Upkeep | per quarter | your estimate | Retrain, retest |
Fill it in with dated numbers from official pricing pages and keep the date beside each.
Compare against not tuning
The OpenAI guide says tuning can reduce prompt length, saving tokens and latency. That saving only counts if your volume is high enough to repay the project. If a better prompt gets you most of the way, you may spend nothing on training. See when to fine-tune or just prompt.
Quick check
What is usually the biggest cost?
Data work and repeated experiments, not the final training run.
Why no price figures here?
Prices change by provider and date. Take current figures from official pages and record the date.
Sources
- Textbooks Are All You Need, fetched 19 October 2026.
- QLoRA: Efficient Finetuning of Quantized LLMs, fetched 19 October 2026.
- LoRA: Low-Rank Adaptation of Large Language Models, fetched 19 October 2026.
- OpenAI model optimization guide, fetched 19 October 2026.
Last reviewed: 19 October 2026.
