Cost formula
Training cost = training tokens / 1,000,000 × training price per 1M; Inference cost = (input tokens / 1M × fine-tuned input price) + (output tokens / 1M × fine-tuned output price)
Calculator
Fine-tuning cost calculator
Project training tokens × epochs, then add monthly inference so you can compare fine-tuning against prompt-only approaches.
Train + serve
Fine-tune beats prompt-only
$12.90train + first month
Ongoing serve $8.40/mo · prompt-only $12.90/mo
Train + serve (mo 1)$12.90
Prompt-only / mo$12.90
Training cost$4.50
Serve / request$0.00042
Prompt-only / request$0.000645
Break-even requests20,000
Workload
Training + inference
Break-even
Train + serve vs prompt-only
Prompt-only path keeps the same model rates but bills the few-shot overhead every request.
Break-even volume20,000 req
Training pays back after about 20,000 served requests (~1.0 months at your volume). Savings $0.000225 per request vs prompt-only.
Inference uses standard list input/output rates for the selected model. Updated 2026-07-31. Approximate static list prices for planning only. Always verify against each provider’s official pricing page before production budgeting.
Guide
How to get value from this calculator
Project one-time training cost and ongoing inference cost for fine-tuned models. Training is billed by token processed; inference uses model-specific rates that may differ from the base model. Plan both phases before committing to custom weights for a use case that a prompt might solve cheaper.
Why it matters
Fine-tuning adds upfront training spend and sometimes higher or separate inference pricing. Without modeling both, teams adopt custom models for problems where a strong system prompt on a cheaper base model would cost less at their monthly volume.
How to use it
- Estimate training dataset size in tokens (examples × average tokens per example × epochs).
- Select the base model and fine-tuning tier offered by your provider.
- Review one-time training cost for your dataset and epoch count.
- Enter expected monthly inference input and output volume post-deployment.
- Compare total first-year cost (training + 12 months inference) against prompt-engineering baseline.
- Check break-even vs prompt-only to see when training pays back against prompt engineering.
- Confirm current fine-tuning and inference rates on the provider's pricing page.
Planning tips
- Start with a subset of data to validate quality before paying for full-dataset training runs.
- Multiple training experiments multiply cost—budget for failed or discarded runs.
- Fine-tuned models may not support all features (tools, JSON mode, caching)—check before designing architecture.
- If monthly inference volume is low, training cost dominates; at high volume, inference dominates.
- If break-even is many months out, stay prompt-only until inference volume justifies training.
- Synthetic data generation for training sets has its own API cost—use the synthetic data calculator if applicable.
Frequently asked questions
Is fine-tuning always cheaper at inference time?
Not necessarily. Some fine-tuned models have dedicated inference rates. Shorter prompts on a fine-tuned model can save tokens, but you must net that against training cost and any rate premium.
How do I estimate training tokens?
Sum tokens across all training examples and multiply by the number of training epochs. Include validation data if your provider bills it as part of the job.
Are prices here official?
No. CentsPerToken uses approximate list prices for planning. Fine-tuning rates change frequently—verify on your provider's site before submitting a training job.
What about hosting a self-trained model?
Self-hosted fine-tuned weights shift cost to GPU infrastructure. Use the local vs cloud comparer for that economics discussion.