Fine-tuning

How it works

You prepare hundreds to thousands of input and output pairs that show the behavior you want, then train the model further on them. The result is a new model version that follows that pattern more consistently with shorter prompts.

Where you can do this has changed. Google offers supervised tuning for Gemini models on its cloud platform (Google Cloud tuning documentation (opens in a new tab)). OpenAI closed its self-serve fine-tuning to new organizations in May 2026, and existing customers lose the ability to start new training jobs in January 2027 (OpenAI deprecations (opens in a new tab)). Open-weight models can be fine-tuned on your own infrastructure, often with a lighter method called LoRA (Hu and colleagues, 2021 (opens in a new tab)) that trains a small set of extra weights.

Fine-tuning teaches behavior. It is a poor way to teach facts. A fine-tuned model still answers from memory, and that memory is out of date the day your price list changes.

How businesses use it

It is worth the cost in a few narrow cases:

• A high-volume classification or extraction task where a smaller fine-tuned model matches a large model's accuracy at a lower price per call
• A strict output format or house style that prompting alone does not hold
• Specialist vocabulary the base model handles poorly
• Lower latency, for example in voice agents, where response time is part of the experience

For company knowledge such as policies, product details and contracts, use retrieval-augmented generation. It costs less, updates the same day and can cite its sources.

Where it goes wrong

Most fine-tuning ideas should start as prompt engineering plus retrieval. Try that first, measure it on a test set, and fine-tune only if a clear gap remains.

The costs that surprise people come after the first training run: assembling clean training data, and redoing the work when the provider retires the base model. A fine-tuned model is tied to the model it came from, so a provider's newer release does not inherit your training.

On a corporate AI consulting engagement, this is one of the decisions we write down with the costs on both sides, so the call is made on numbers.