Fine-tuning
What is Fine-tuning?
Taking a pretrained model and training it further on your own data so it adapts to a specific task, domain or style. It updates the model's parameters, unlike RAG, which supplies knowledge at query time. Fine-tuning suits fixed patterns and a consistent tone. For facts that change, retrieval usually wins, and the two are often combined.
Why it matters
Fine-tuning is a commitment, and pricing it honestly changes when it makes sense. The visible cost is the training run, but the real bill includes preparing and cleaning the examples, evaluating whether the result is actually better, and redoing the whole exercise when the base model you built on is superseded, which happens often. That expense pays off for a narrow, stable job: a fixed output format, a consistent house style, a specific classification the model must get right every time. It rarely pays off for knowledge that changes, where feeding fresh information at query time is cheaper and stays current. The counter-intuitive win is size. A small model tuned for one task can beat a much larger general model on that task, and cost far less to run, which matters enormously once it is answering millions of times a day.
In practice
A company needs every response in a strict format its downstream systems can parse. Rather than prompt a large general model and validate the shape each time, it fine-tunes a small model to produce that format natively, which runs faster and cheaper at the volumes it handles. The catch is maintenance. When the base model gets a major update, the tuning has to be evaluated and often redone, so the saving is booked against that recurring cost.
