What is fine-tuning an AI model?
Training an existing model a little further on your own examples so it picks up a task, a format or a style. Often a good prompt does the job first.
intermediate · Also called Fine Tuning, Finetuning
Every assistant you use has already been fine-tuned. A model fresh from pretraining is good at continuing text, not at following instructions. It becomes ChatGPT or Claude through more training on examples of good answers. Fine-tuning is the same move made by a customer: take a finished model and train it a little more, on your own examples, so it behaves the way you need.
What it is good for
Fine-tuning teaches how, more than what. It works well for:
- A fixed format. Always reply in the exact structure your system expects.
- A narrow task. Sorting support tickets into your 12 categories.
- A house style. The tone your brand writes in.
- A smaller model. Teaching a cheap model one job well enough that you do not need an expensive one.
OpenAI lists the benefits as handling more examples than fit in a prompt, needing shorter prompts and making a smaller, cheaper model viable.
What it is not good for
Keeping a model up to date with facts. Fine-tuning makes a model mimic the patterns in its examples, and a fact that changes next month would mean training again. For knowledge that changes, give the model the documents when it answers, with RAGRAGA way to answer from your own documents. A search finds the passages that match the question, and they are added to the prompt before the model writes. or search.
Try the prompt first
OpenAI tells developers to write tests first, because a better prompt may be all you need. A few good examples in the prompt and a clear system promptSystem PromptInstructions the builder of an AI app gives the model before you type anything. It sets the role, the tone and the rules, and it outranks what users say. are quicker to try, and cheaper to change.
The ground is shifting
Fine-tuning is getting harder to buy from the largest providers. Anthropic's API does not offer it. OpenAI announced in May 2026 that it is winding down its fine-tuning platform: new users were shut out at once, and active existing customers can no longer start new training jobs from 6 January 2027. Models already fine-tuned keep working until the model they were built on is retired.
Open models are another matter. Whoever downloads one can train it further on their own hardware, and methods such as LoRALoRALow‑Rank Adaptation, a parameter‑efficient fine‑tuning method exist to make that cheaper.
Before you start
Collect real examples of the input and the output you want. Write down how you will judge success. Then see how far a careful prompt gets you. If it falls short on format or consistency, not on facts, fine-tuning may be worth it.
Sources
Where each fact comes from
This page was last checked against these sources on 14 October 2026.
- 1Glossary
Anthropic, Claude API documentation · 14 October 2026
Fine-tuning is further training of a pretrained language model with additional data, so the model starts to mimic the patterns of the fine-tuning dataset. Claude has already been fine-tuned to be a helpful assistant. The Claude API does not currently offer fine-tuning. Fine-tuning can adapt a model to a domain, task or writing style, and needs careful thought about the data and its effect on performance and biases.
- 2Model optimization
OpenAI API documentation · 14 October 2026
Fine-tuning takes a base model and the kinds of inputs and outputs you expect and gives a model that excels at those tasks. It can handle more examples than fit in a prompt, allow shorter prompts and a smaller, cheaper model. Supervised fine-tuning trains on correct responses, DPO on correct and incorrect ones, reinforcement fine-tuning grades outputs, and vision fine-tuning uses images. Write evals first, because prompt engineering may be all you need. OpenAI is winding down the fine-tuning platform. It is no longer open to new users, existing users can create training jobs for the coming months, and fine-tuned models stay available until their base models are deprecated.
- 3Deprecations
OpenAI API documentation · 14 October 2026
Developers were notified on 7 May 2026. From 2 July 2026 organisations that had not run inference on a fine-tuned model in the past 60 days can no longer create fine-tuning jobs. From 6 January 2027 active existing customers can no longer create new fine-tuning jobs. Inference on fine-tuned models is disabled only when the underlying base model is deprecated.
- 4RAG (retrieval augmented generation), in the Glossary
Anthropic, Claude API documentation · 14 October 2026
RAG passes documents into the context window at runtime, so the model can use information beyond its training data. It is particularly useful for tasks that need up-to-date information.
- 5How ChatGPT and our foundation models are developed
OpenAI Help Center · 25 September 2026
Models are developed in stages that include pre-training and post-training.