What Is Fine-Tuning? RAG, Data Requirements, Pricing, and Availability

“I want to customize AI for my company.” One option is fine-tuning: additional training of a pretrained LLM on data tailored to a particular use. Data requirements, costs, and available services vary, so first establish a baseline for measuring improvement. This beginner’s guide explains what fine-tuning does, its strengths, how it compares with RAG and prompting, the main methods, what you need, and how to get started.

1. What is fine-tuning?

Fine-tuning means training an already trained AI model (the base model) further on data for your use case to adjust its response patterns. Examples include answering in your company’s style, using a particular output format, or adapting to specialist terminology. The aim is to teach these patterns to the model itself.

Think of it as new employee training. Even a capable recruit—the base model—does not know your company’s practices. Training on company examples aims to adapt responses to those practices and lets you evaluate whether shorter instructions will suffice. Fine-tuning updates model weights (parameters) or additional adapters.

A “custom model” does not mean you own the underlying model. Distinguish a model accessed through an API from one whose weights you handle yourself, and check the base model’s license and distribution terms.

2. Strengths and limitations

Misunderstanding the purpose can lead to failure. Fine-tuning is an option for adjusting response patterns such as style, classification, and instruction following; it is not a mechanism for automatically incorporating information that appears after training.

If the goal is to handle current information or internal data accurately, RAG (retrieving information and adding it to context) is a better fit than fine-tuning. Fine-tuning is an option for bringing responses closer to a desired tone or format. If you need a format such as JSON, also compare structured outputs on supported models. OpenAI’s structured outputs guide distinguishes format compliance from the semantic correctness of values and explains how to handle refusals.

3. Choosing between RAG, prompting, and fine-tuning

These three common approaches to customizing AI have different roles and cost drivers. The table gives an overview.

A useful rule of thumb is: “RAG for facts and knowledge; fine-tuning for behavior and patterns; prompts first.” In production, you can combine RAG to supply documents with fine-tuning to adjust responses, but not every system needs both. This connects with context engineering.

4. Main methods: full fine-tuning, LoRA, and QLoRA

SFT learns from desired outputs. DPO uses response preferences, while RFT uses signals such as grader scores. Full fine-tuning, LoRA, and QLoRA below are a different classification: they concern which parameters are updated and numerical precision. API services do not necessarily let you choose the internal implementation freely. See OpenAI’s training methods guide.

There are several ways to fine-tune. Beginners should first understand these three.

The official PEFT documentation and the original LoRA/QLoRA papers explain the mechanisms. Experimental results in those papers are not guarantees for every model.

Choose full fine-tuning, LoRA, or QLoRA based on available resources and evaluation results. Failure with one method does not establish that the others will fail. With a local LLM, a small trial on your own PC may also be possible.

5. Data, costs, and tools

Data requirements: an accepted minimum is not the amount needed for useful quality

OpenAI’s SFT guide reports improvements with 50–100 examples and recommends revisiting the task or prompt if 50 examples produce no improvement. This is guidance for OpenAI’s SFT, not evidence that every method requires 500 examples. Meeting the submission minimum and achieving useful quality are separate questions.

Before increasing the count, check whether examples resemble real inputs, answers are correct, classes are balanced appropriately, and duplicates or contradictions exist. Separate training and evaluation data so that paraphrases of the same question do not appear in both and inflate scores. Teaching a correct answer when the input lacks the information needed to derive it may encourage unsupported guesses. See OpenAI’s guidance on data quality and separating evaluation data.

Pricing: include ongoing use, not just training

For token billing, start with “billable training tokens ÷ one million × unit price”; for time billing, use “billable time × hourly rate.” Check which tokens or time count and how failures and retries are treated by your provider. For self-managed training, also estimate passes through the data, input length, GPU count, and running time.

These two examples use different methods and units, taken from the Standard rates in OpenAI’s official pricing table, checked on October 10, 2026. Their inclusion does not mean new users can sign up; see the availability conditions below.

o4-mini and its fine-tuned version in this pricing example are scheduled to shut down on October 23, 2026. Check OpenAI’s official deprecation notice before choosing a model.

These are API fine-tuning and inference prices, not ChatGPT subscription prices. Batch processing and data sharing may have separate rates, so “$25–$100 per million tokens” is not a price shared by every method. Include sharing conditions and inference rates in your estimate.

If the aim is to save money, measure whether the cost per request actually falls after training. Dividing implementation costs by the saving per request gives an approximate break-even volume, but only when response quality is comparable and the saving is positive. For low-volume tasks, reduced inference costs may not recover preparation costs.

Availability: OpenAI’s phase-out and self-managed training

Source: OpenAI’s self-service fine-tuning availability changes, checked on October 10, 2026. A pricing table or how-to guide remaining online is not evidence that your organization can start.

For self-managed training with a model whose weights you can access, Hugging Face TRL’s SFTTrainer is one option. Check model and data licenses, GPU memory, input length, supported training methods, and storage destinations. For another provider’s managed service, likewise verify supported models, regions, contract terms, and billing units against its current specifications.

6. A small trial to decide whether to adopt it

The following is an illustrative plan for classifying inquiries as “shipping,” “returns,” or “other.” It is not a report of training runs, measured performance, or bills from this article.

1. Define acceptance criteria: record classification accuracy, handling of uncertain cases, response time, and cost per request. Track serious misclassifications separately from average performance.
2. Measure baseline performance: organize instructions and examples and evaluate on data that will not be used for training. Distinguish missing instructions, missing information, and model judgment.
3. Choose a method that addresses the cause: if reference material is missing, review the retrieval system. If decision criteria or response patterns are the issue, include fine-tuning in the comparison. You do not always need to try RAG first.
4. Check execution conditions: verify availability, model, budget, data usage permissions, external transmission, and storage before planning a small training run.
5. Compare using the same evaluation: give the same questions to the base and fine-tuned models. Include unseen inputs, other languages, long text, and requests that should be refused.
6. Decide whether to adopt, adjust, or defer: compare quality improvements, costs, latency, and operational workload. If there is no improvement, revisit data or settings rather than immediately replacing production. Keep a version you can revert to.

After training, continue monitoring areas where errors increased, information updates, and base-model deprecations. Training performance alone does not show whether the model handles unseen inputs. OpenAI’s evaluation guidance also recommends separating training and test data.

Summary

Three points to remember about fine-tuning:

• What it is: additional training of a pretrained model to adjust response patterns by updating weights or additional adapters.
• How to choose: identify whether the problem concerns reference information, instructions, or response patterns, then compare methods under the same conditions.
• How to start: check availability, budget, and data usage permissions. Data requirements and costs vary by method, model, and use case.

Fine-tuning is an improvement option whose effectiveness should be established through evaluation. Completing a training job and improving quality in real work require separate checks. Related guides to RAG and context engineering help explain the broader picture of AI customization. Another related technique for making models smaller is model distillation.

FAQ