LLM Fine-Tuning Cost Estimator

Fine-tuning cost = (total training tokens) × (price per 1K tokens). Total training tokens = examples × avg tokens per example × epochs. For example, 1,000 examples × 500 tokens × 3 epochs = 1.5M tokens. At OpenAI GPT-3.5 Turbo fine-tuning rate of $0.008/1K tokens, that's $12 for training. Add inference costs separately at the fine-tuned model rate.

The default of 3 epochs works well for most datasets. More epochs improve fit on small datasets but risk overfitting. If your dataset is large (>10K examples), 1–2 epochs often suffice. OpenAI recommends starting with 3 epochs and adjusting based on validation loss. Each epoch multiplies your training cost linearly.

OpenAI requires at least 10 examples but recommends 50–100+ for meaningful improvements. For specialized domains, 200–1,000 high-quality examples typically outperform thousands of mediocre ones. Anthropic and Together AI have similar minimums. Quality matters far more than quantity — curate carefully rather than scale blindly.

Fine-tuning makes sense when: (1) few-shot prompting still produces inconsistent output, (2) you need a specific format or tone that's hard to describe in a system prompt, (3) you're making 100K+ inferences/month and the cost of a smaller fine-tuned model beats repeated few-shot prompting, or (4) you need to compress a complex 4K-token prompt into the model's weights. Don't fine-tune to add new knowledge — use RAG for that.

No. Fine-tuning teaches format, style, and behavior — not new factual knowledge. The model's training cutoff doesn't change. For updating knowledge, use RAG (Retrieval-Augmented Generation) to inject context at inference time. Fine-tuning is for teaching *how* to respond, not *what* to know.