Fine-Tuning an LLM: Cost and When It Is Worth It
your project?
Apple just put an LLM-based Siri on the iPhone 18 Pro, running partly on-device on its new A20 chip. That is a preview of where a lot of AI products are heading: smaller, specialized models doing one job well instead of a general-purpose model doing everything. Fine-tuning is how you get there, and the cost of doing it ranges from under $10 to well over $10,000 depending on choices most teams do not weigh carefully enough.
Apple's push toward on-device, LLM-based Siri on the iPhone 18 Pro is a visible signal of a broader shift: general-purpose cloud models are not the only answer anymore, and smaller, purpose-built models fine-tuned for a specific task are becoming a real competitive option. LLM fine-tuning cost varies enormously depending on model size and method, and knowing where you land on that range before you start is the difference between a cheap, fast win and an expensive, failed experiment.
The real cost range for 2026
- LoRA or QLoRA on a small model (2–8B parameters): often under $10 on cloud GPUs for a single run, and realistically $40 to $1,600 once real data and iteration are included;
- LoRA on a 70B model: can run under $500 on consumer-grade hardware, since parameter-efficient methods scale far better than full fine-tuning;
- Full fine-tuning of a 7B model: typically $1,000 to $12,000, depending on the infrastructure provider and training run count;
- Full fine-tuning of larger models: can exceed $10,000 per run, before accounting for failed attempts and iteration.
Training compute itself varies from roughly $0.48 per million tokens for open-source 7B models to $25 per million tokens for a frontier model like GPT-4o, which is a 50x spread before you have picked anything else about the project.
The hidden costs that double the estimate
The training run itself is rarely the number that breaks a budget. What actually doubles it:
- data preparation and labeling, which can dwarf the compute cost. One documented case spent $12,000 on labeling 500 legal documents with three reviewers each, more than the training itself;
- failed training runs, which are normal, not exceptional. One team spent $8,000 across two failed runs before a usable model emerged;
- storage and versioning, for datasets, checkpoints, and every fine-tuned variant you keep around;
- ongoing evaluation and re-tuning, since a fine-tuned model needs the same evaluation discipline as any other model in production, on a recurring basis, not once.
A quote for "fine-tuning" that only covers the training run is quoting the cheapest part of the project. Data preparation and iteration are usually where the real budget goes.
Cutting the cost without cutting corners
Spot and preemptible GPU instances are the single biggest lever available, cutting training cost by 60 to 70 percent compared to on-demand pricing. A full fine-tune priced at roughly $1,638 on-demand can run closer to $490 on spot instances, for the same training job. The trade-off is tolerance for interruption, which matters for long runs but rarely for the shorter LoRA-based ones most teams actually need.
When fine-tuning is actually worth it
Fine-tuning is not always the right tool, and the fastest way to overspend is choosing it by default.
- choose fine-tuning when you need consistent tone, format, or behavior on a narrow, repeatable task, and the underlying knowledge is stable rather than changing weekly;
- choose RAG instead when the problem is your AI not knowing your current facts or documents, since retrieval solves that more cheaply and without a training run at all;
- consider on-device or a small fine-tuned model when latency, privacy, or cost per inference matters more than raw capability, echoing exactly the direction Apple is pushing with an LLM-based Siri running partly on-device;
- do not fine-tune to patch a data problem. If the model is missing facts, more training data of the same stale information will not fix it. That is what retrieval is for.
Questions to ask before you commit budget
- Is this a knowledge problem or a behavior problem? Fine-tuning fixes behavior; it rarely fixes missing or changing facts.
- What method fits the model size? LoRA for most practical cases; full fine-tuning only when the task genuinely demands it.
- Is the data prep cost included in the quote, separately from the training compute?
- What is the plan for failed runs? Budgeting for one successful attempt on the first try is optimistic, not realistic.
- Can this run on spot instances? If the job tolerates interruption, this is free money left on the table by not asking.
Conclusions
LLM fine-tuning cost ranges from under $10 for a small LoRA run to well over $10,000 for a full fine-tune on a large model, and the real number is set far more by data preparation, failed runs, and ongoing evaluation than by the training compute itself. As Apple pushes on-device, purpose-built AI with an LLM-based Siri on the iPhone 18 Pro, the case for smaller, fine-tuned models over one giant general-purpose model keeps getting stronger. Confirm you actually have a behavior problem before you fine-tune, budget for the hidden costs, and use spot instances wherever the job tolerates it.