AI concepts
Fine-tuning (tuning an AI model to a specific task)
The process of adapting a pre-trained LLM to a specific task by further training it on dedicated data. Fine-tuning costs 5 to 50 times more than prompt engineering and 2 to 10 times more than RAG, but it delivers better results in three narrow classes of use case: replicating a style, narrowing scope, and reducing inference costs at high volume.
Primary source: OpenAI Fine-tuning Best Practices 2025, Anthropic Constitutional AI 2024
The first decade of NLP (2010 to 2020) relied on fine-tuning as the main method of adapting models. In the LLM era (2022 onward) its role is diminishing. In 2025, fine-tuning is a niche tool, used in fewer than 8 percent of enterprise GenAI deployments (Gartner estimates, Q1 2026).
When fine-tuning makes sense
Replicating an organisation's style. A bank wants its chatbot to write offers in the same tone as the product department. Prompt engineering reaches 70 to 80 percent similarity. Fine-tuning on 5,000 historical offers reaches 95 to 99 percent. A brand-critical difference.
Narrowing scope. A model fine-tuned for radiology diagnostics hallucinates less often than GPT-4 with RAG, because it has no broad "general" knowledge to mix in.
Cost reduction at scale. A fine-tuned smaller model (Llama 3.1 8B) can replace GPT-4 for a specific task at 1/30 of the inference cost. It pays off above 10 million queries a month.
When fine-tuning does not make sense
Knowledge that changes over time. Every data change requires retraining (from a few hours to a few days). RAG updates instantly.
A small dataset. Fine-tuning on fewer than 1,000 examples gives worse results than a good prompt. Anthropic recommends a minimum of 500 to 1,000 examples per task.
Factual questions. RAG is cheaper and more precise for "what did client X write in the March 2024 report".
Operational cost
OpenAI fine-tuning of GPT-4o-mini (Q1 2026): around 25 dollars per million tokens of training, 0.30 dollars per million tokens of inference (vs 0.15 for non-fine-tuned). For a base of 5,000 examples, each 1,000 tokens, the training cost is around 125 dollars. Modest. The operational cost: double that of the base model.
Open-weights models (Llama, Mistral) fine-tuned on your own infrastructure: the cost of hardware (1 to 4× A100 GPU = 30 to 120 thousand dollars, or the cloud equivalent) plus the operational cost of the GPU plus the cost of an ML team maintaining the model.
Polish context
In Polish firms in 2025 and 2026, fine-tuning was used mainly by the financial sector (3 banks from the top 10), the legal sector (2 law firms), and large e-commerce firms. The total population of deployments: fewer than 30 cases. The dominant pattern is prompt engineering plus RAG, not fine-tuning.
The decision whether fine-tuning makes sense for a specific use case is part of the AI Readiness Audit.