Fine-Tuning Cost Calculator

Estimate training costs and ongoing inference pricing to decide whether fine-tuning makes financial sense for your use case.

Training & inference prices per 1K tokens (approx.).
Total words across all examples.
How many passes over the data.
For inference with the fine-tuned model.
Typical completion length.
For ongoing inference cost.

Frequently asked questions

Is fine-tuning worth it vs. few-shot prompting?

Fine-tuning makes sense when: (1) you need consistent output formatting that few-shot examples can't reliably achieve, (2) you're running millions of requests/month and a shorter fine-tuned prompt saves more than the training cost, or (3) you need domain knowledge not in the model's training data. For most cases under 100K requests/month, a well-crafted prompt is cheaper.

How many training tokens do I need?

Effective fine-tuning typically uses 50–500 high-quality prompt-completion pairs. At ~250 tokens per pair, that's 12,500–125,000 tokens per epoch. Running 3 epochs means 37,500–375,000 training tokens total. Quality matters far more than quantity — 100 excellent examples outperform 10,000 mediocre ones.

How much does fine-tuning GPT-5 Mini cost?

Training GPT-5 Mini costs approximately $5/1M training tokens. A small run with 100 examples × 300 tokens × 3 epochs = 90,000 training tokens costs roughly $0.45. The fine-tuned model then uses the same per-token inference pricing as the base model. The main saving comes from shorter prompts (no few-shot examples needed).

When does fine-tuning break even vs. prompting?

Calculate your per-request savings from a shorter prompt (fewer few-shot examples), then divide the training cost by that saving. If your training cost is $5 and each request saves $0.001 in prompt tokens, you break even after 5,000 requests. At 50,000 requests/month, you'd recoup costs in the first week.