How fine-tuning works
A base model already knows language and a great deal about the world. Fine-tuning runs a short additional round of training on your own examples — pairs of inputs and the outputs you want — so the model's weights shift toward your task. The result is a new model version, served under its own identifier, that produces your format or style with a shorter prompt than the base model would need.
It is a different tool from retrieval. Fine-tuning changes how a model behaves; it is a poor way to load facts that change, because updating them means another training run. RAG changes what the model sees at query time and keeps facts current. Many teams that reach for fine-tuning to add knowledge are better served by retrieval.
Where fine-tuning is offered in 2026
The major first-party APIs have pulled back. OpenAI's pricing page states that it is winding down its fine-tuning platform: it is no longer accessible to new users, existing users can create training jobs for the coming months, and fine-tuned models remain available for inference until their base models are deprecated. Google's Gemini API documentation says fine-tuning is not available there — no model supports it since Gemini 1.5 Flash-001 was deprecated — with tuning offered on its enterprise platform instead. Mistral's documentation marks its fine-tuning feature as deprecated and no longer actively supported.
That leaves open-weight models as the main route: their weights can be fine-tuned on your own hardware or through hosting providers, and the tuned copy is yours to keep running.
What fine-tuning costs
There are three parts. Training: OpenAI's remaining fine-tuning offer bills training by the hour, while other providers have billed per training token; either way, cost scales with dataset size and the number of passes. Inference: a tuned model is billed per token like any other, but often at its own rate — OpenAI lists separate input and output rates for its fine-tuned model and discounts them if you enable data sharing. Upkeep: some providers charge monthly to store a tuned model, and every base-model retirement means retraining on a new base.
The saving has to come from shorter prompts or a smaller model doing the job of a larger one. If a tuned small model replaces a large model with a long prompt, inference savings can repay the training; if the prompt barely shrinks, they rarely do.
Try these first
- Better prompts with examples — a few good examples in the system prompt, cached, often match a tuned model for format and tone.
- RAG — for knowledge, retrieval beats tuning on freshness and citations.
- A different model — the current generation may already do what the old one needed tuning for.
- Routing — send the repetitive easy cases to a small model and keep the large one for the rest.