How to use the Fine-Tuning Cost Calculator
- 1. Size your dataEnter the number of examples, average tokens per example and epochs.
- 2. Choose how to trainUse a provider's price per million training tokens, or your measured GPU throughput and hourly price.
- 3. See the estimateGet the training cost, and optionally the monthly cost of using the tuned model.
The formula
Training cost depends on how many tokens the model processes. Each epoch is one full pass over your data, so:
training tokens = examples × average tokens per example × epochs
With a provider API, cost = training tokens ÷ 1,000,000 × price per million. On your own GPUs, time = training tokens ÷ throughput (tokens per second), and cost = hours × your hourly price. Multiply by the number of runs you expect, because the first run is rarely the last.
Worked example
The defaults are 2,000 examples of about 600 tokens each, trained for 3 epochs, which is 3.6 million training tokens. At $5 per million (OpenAI's listed gpt-4.1-mini fine-tuning price on 7 November 2026), one run costs $18, about ₹1,692 at ₹94 per dollar. Tick the box for monthly use: 50,000 requests a month with 800 input and 200 output tokens, at $0.80 and $3.20 per million, add about $64 (₹6,016) a month.
Before you fine-tune
- Try better prompts and retrieval (RAG) first. They're cheaper and often enough.
- OpenAI says it is winding down its fine-tuning platform and it's closed to new users, so check what your provider offers today.
- For your own GPUs we don't pre-fill throughput or prices. They vary hugely by model size, GPU, LoRA vs full fine-tuning and provider, so measure a short run.
- Budget for evaluation runs and failed experiments, not just the final training job.
Frequently asked questions
Are tokens counted the same everywhere?
No. Each provider uses its own tokenizer, so the same file can be a slightly different number of tokens. Use the provider's tools on your file for an exact count.
What's a reasonable number of epochs?
Commonly 1–4 for instruction tuning; more can overfit small datasets. Watch validation loss rather than picking a number up front.
Is the fine-tuning cost calculator free?
Yes. You can open the fine-tuning cost calculator and fill in everything without signing up. When you estimate a cost for the first time, we ask once for your name and WhatsApp number; after that it, and every other Uncodemy tool, opens straight away.