Fine-tuning
Last verified
Fine-tuning adapts a base model to your data. It has two costs: the training run itself, and typically a higher per-token inference rate for the resulting model.
That second cost is the one people forget. A fine-tune that runs on every request can be more expensive in month two than the training was in month one.
Compare against the alternatives first — a better prompt, few-shot examples, or retrieval — which have no training cost and no rate premium.