Fine-tuning

Fine-tuning is the process of taking a pre-trained language model and continuing to train it on a smaller, more specialised dataset. Instead of training a model from scratch (which requires billions of dollars of compute), fine-tuning adjusts an existing model’s weights to specialise it for a specific domain or behaviour.

In 2026, fine-tuning is available for some models (OpenAI offers it broadly; Anthropic offers it for select enterprise customers; open-source models like Llama support it freely). The Claude.ai and Claude Code consumer products don’t expose fine-tuning to end users.

When fine-tuning helps

• You have a lot of well-labelled examples (hundreds to thousands) of the exact behaviour you want
• The behaviour is hard to specify with a prompt (e.g. matching a very specific writing style, classifying highly nuanced edge cases)
• You want to reduce token costs at inference time by baking instructions into the model

When fine-tuning doesn’t help (and you should just write a better prompt)

• Your data is small (under ~100 examples)
• The behaviour is easy to describe (“respond in Australian English”)
• You need the model to know new facts (use RAG, not fine-tuning, facts go stale)
• You haven’t tried prompting hard yet

The “fine-tune first” instinct is usually wrong in 2026. Modern frontier models follow instructions well enough that prompt engineering + RAG covers most use cases more cheaply and with more flexibility.

Fine-tuning vs RAG vs prompting

Approach	Best for	Cost
Prompting	Most tasks	Per-call inference only
RAG	Grounded answers from your data	Inference + vector DB
Fine-tuning	Style mimicry, classification, narrow workflows	One-time training cost + per-call inference

Australian SMB perspective

In 5+ years of running AI consulting work, we’ve recommended fine-tuning to a client maybe twice, both times for very narrow, high-volume classification tasks. For 99% of Australian small business AI use cases, prompt engineering + RAG + tool use beats fine-tuning on cost, flexibility, and time-to-value.

If a vendor is selling you fine-tuning as the answer to a generic “we’ll make AI work for your business” pitch: ask hard questions about why a tighter prompt wouldn’t do the same thing for one-hundredth the cost.

Related terms

• Large Language Model
• Prompt
• Retrieval-Augmented Generation