The short answer
Fine-tuning is further training of an already trained AI model on your own examples, so it adopts a particular task, format, style or domain. Google's Machine Learning Glossary calls it a second, task-specific training pass on a pre-trained model. In 2026 it is harder to get from the big providers: OpenAI says it is winding down its fine-tuning platform and no longer accepts new users, and Anthropic says the Claude API does not currently offer fine-tuning. Open-weight models can still be fine-tuned.
What fine-tuning is, in plain English
A large language model is first trained on a huge amount of general text. Fine-tuning is a second, smaller round of training on examples you choose, such as a few hundred support tickets with ideal replies. The model's internal settings shift so that it imitates those examples without being told each time.
Google's Machine Learning Glossary defines fine-tuning as a second, task-specific training pass performed on a pre-trained model to refine its parameters for a specific use case. Anthropic's glossary describes it as further training a pretrained language model with additional data, which makes the model start representing and mimicking the patterns of that dataset. Anthropic adds that it can help adapt a model to a domain, task or writing style, but needs careful thought about the data and its effect on performance and bias.
The assistants you use are already fine-tuned. Anthropic notes that Claude is not a bare language model: it has been fine-tuned to be a helpful assistant, partly through reinforcement learning from human feedback.
Fine-tuning vs prompting vs RAG
| Approach | What changes | Best for | Effort |
|---|---|---|---|
| Prompting | Only the instructions you send | Almost everything, first | Minutes |
| Saved instructions (skills, projects, custom instructions) | Instructions sent automatically each time | Repeated tasks with a house style | An hour |
| RAG | The documents the model can see for each question | Facts that change or must be cited | Hours to days |
| Fine-tuning | The model itself | A narrow task at high volume with a fixed format | Days to weeks, plus data work |
OpenAI's model optimisation guide frames fine-tuning as something used together with prompt engineering, not instead of it. It lists the benefits: training on more examples than fit in one request's context window, shorter prompts that cut token costs and latency at scale, training on sensitive data without sending it every time, and getting a smaller, cheaper model to do a specific task well.
What changed in 2026
- OpenAI is winding down its fine-tuning platform. OpenAI's guide says the platform is no longer accessible to new users, existing users can create training jobs for the coming months, and fine-tuned models stay available for inference until their base models are deprecated. The methods it has offered are supervised fine-tuning, vision fine-tuning, direct preference optimisation and reinforcement fine-tuning.
- Anthropic does not offer it in the Claude API. Its glossary says the Claude API does not currently offer fine-tuning and points interested customers to their Anthropic contact.
- Open-weight models keep it possible. When a model's weights are published, you can train them yourself. DeepSeek publishes DeepSeek-V4-Pro under the MIT licence, Cohere released Command A+ on 20 May 2026 with open weights under Apache 2.0, and Mistral AI's Mistral Medium 3.5 is a 128B open-weight model. See what is an open-source model and open-source ChatGPT alternatives.
Why fine-tuning matters at work
For most office teams the useful lesson is negative: you probably do not need it. If your goal is "make the AI write like us", a saved skill or project with three good examples gets you most of the way in an afternoon. If your goal is "make the AI know our policies", that is a RAG problem, because policies change and answers should cite them.
Fine-tuning earns its cost when all of these are true: the task is narrow and repeated thousands of times, the output format must be identical every time, you have hundreds of high-quality examples, and prompting plus examples has already been tried and falls short. Typical cases are classifying tickets into a fixed set of categories or extracting the same fields from a standard form, usually built by an engineering team through an API.
The question also matters for privacy. When a vendor says it does not "train or fine-tune" on your data, it is promising that your prompts will not be used to change its models. Make, for example, says prompts and outputs in its Maia assistant are not used to train or fine-tune the underlying model. Check the equivalent setting in each tool; see how to keep your data private in AI tools.
Try it: get fine-tuning results without fine-tuning
- Collect five examples of the output you want, such as five product descriptions your team is proud of.
- Open a new chat and paste the prompt below with the examples.
- Generate three new pieces and compare them to your examples side by side.
- When the results are right, save the instruction and examples as a skill in ChatGPT, Gemini or Copilot, or as project instructions in Claude. See how to set up Claude projects and how to make AI sound like you.
If the saved version is good enough, and it usually is, you have avoided weeks of work.
Common misconceptions about fine-tuning
- "Fine-tuning teaches the model our facts." It shapes behaviour and style far more reliably than it stores facts. For facts, use RAG.
- "Uploading files to ChatGPT fine-tunes it." No. Uploads and projects add material to the context or retrieve from it. The model is unchanged.
- "A fine-tuned model will not hallucinate on our topic." It can still produce confident errors; see what is hallucination.
- "Fine-tuning is the professional choice." Anthropic's agent guidance applies here too: start with the simplest solution that works.
- "Once fine-tuned, it is ours forever." OpenAI says fine-tuned models stay available only until their base models are deprecated.
Related terms
- RAG is the usual better choice for company knowledge.
- Open-source model covers models whose weights you can fine-tune yourself.
- System prompt means standing instructions, the lightweight alternative.
- Custom GPT is a saved assistant, not a fine-tuned model.
- AI API is where fine-tuning is offered, when it is.
Prompts to copy
Few-shot style match instead of fine-tuning
Here are five examples of [product descriptions / support replies / summaries] in our house style: [paste examples]. First, describe the style in five rules (length, structure, tone, words we use, words we avoid). Then write a new one for: [details]. Follow the rules and match the examples.
You get: A written style guide plus a new piece that matches your examples, ready to save as a skill or project instruction.
Decide if you need fine-tuning
We want AI to [goal]. Volume: [how often]. We have [number] examples. Ask me questions to decide whether prompting, saved instructions, RAG or fine-tuning fits best, then recommend one with reasons and the first three steps.
You get: A reasoned recommendation for the lightest approach that meets your goal.