In this article
Example: a loan rate changes
With fine-tuning: 2 to 4 days
- 1Edit examplesWith the new rate: half a day
- 2Train againAnd re-test: 1 to 3 days
- 3Ship the new modelHalf a day
With RAG: minutes
- 1Upload documentCurrent policy: 5 min
- 2Update the searchOnly that document: minutes
- 3AnswerCiting the section
Which one to choose
| Criterion | RAG | Fine-tuning |
|---|---|---|
| What changes | What the model is given on each query | The model itself |
| Updating a fact | Reload the document: minutes | Train and test again: days |
| Cites the source | Yes, document and section | No |
| Per-user permissions | Yes, per document | No: anyone sees what it learned |
| Deleting a customer's data | Delete the document | You must train again |
| Price per token | Current model gpt-6-luna (USD 0.10 / 0.50 per million), with more text per query | Fine-tuned gpt-4.1-mini (USD 0.80 / 3.20 per million): 8 times more per input token |
| What you need | Current, organized documents | Reviewed input and output examples |
| Best for | Policies, rates, contracts, manuals | Fixed format, tone, classification |
What changes
- RAG
- What the model is given on each query
- Fine-tuning
- The model itself
Updating a fact
- RAG
- Reload the document: minutes
- Fine-tuning
- Train and test again: days
Cites the source
- RAG
- Yes, document and section
- Fine-tuning
- No
Per-user permissions
- RAG
- Yes, per document
- Fine-tuning
- No: anyone sees what it learned
Deleting a customer's data
- RAG
- Delete the document
- Fine-tuning
- You must train again
Price per token
- RAG
- Current model gpt-6-luna (USD 0.10 / 0.50 per million), with more text per query
- Fine-tuning
- Fine-tuned gpt-4.1-mini (USD 0.80 / 3.20 per million): 8 times more per input token
What you need
- RAG
- Current, organized documents
- Fine-tuning
- Reviewed input and output examples
Best for
- RAG
- Policies, rates, contracts, manuals
- Fine-tuning
- Fixed format, tone, classification
Example: a bank wants to answer loan requirements (they change every month) in the complaints team's format (fixed). Requirements: RAG. Format: prompt instructions; fine-tuning only if quality tests show they are not enough.
In Paraguay and Ecuador, your customers can ask you to delete their data. With RAG you delete a document; with fine-tuning, that data stays in the model itself. Do not train on personal data.
- Paraguay: right to erasure, Law 7593/2025, art. 31.
- Ecuador: right to deletion, LOPDP, Registro Oficial 459.
- Banks: citing the source of each answer helps with the BCP's SIB (PY) or the Superintendencia de Bancos (EC).
- By country: AI for banks and fintech in Paraguay · AI for banking and cooperatives in Ecuador.
What changed in 2026
2027-01-06
is the last day active OpenAI customers can create fine-tuning jobs. Since 2026-05-07 it does not accept new organizations. Already tuned models stay live until their base model is retired.
Criterion: before investing in fine-tuning, confirm your provider offers it for that model and until when.
What each option costs per month
RAG or fine-tuning: total cost per month
- RAGcheaperUSD 76.95Model gpt-6-luna USD 1.95 + search USD 50.00 + team USD 25.00
- Fine-tuningUSD 430Fine-tuned gpt-4.1-mini USD 6.00 + training USD 24.00 + team USD 400
With content that changes, RAG is cheaper: updating a document takes minutes and nothing has to be trained again.
Tokens: the unit AI providers bill for the text the model reads (input) and writes (output). OpenAI puts 100 tokens at about 75 English words.
Prices from OpenAI API pricing (2026-10-11). RAG on gpt-6-luna: USD 0.10 input and USD 0.50 output per million tokens. Fine-tuning on fine-tuned gpt-4.1-mini (fine-tuning only exists on older models): USD 0.80 and USD 3.20; training USD 5.00 per million tokens.
Assumptions: RAG uses 4,000 input tokens (question and document excerpts) and 500 output; the tuned model, 500 and 500. Examples of 800 tokens, 3 passes per training run. With RAG each change takes 1 team hour; with fine-tuning, the hours you choose (16 at the start, about 2 days). Search server, hours, technical team hourly cost, minutes saved and setup are examples. The hourly cost of the person who does the work today is anchored to the minimum wage of Paraguay (MTESS) or Ecuador (SBU ÷ 173.33 h); the starting value is Ecuador's example real cost, USD 6.95. Pays back in = setup ÷ (value of the hours saved − monthly total). With the starting values: RAG 1.95 + 50 + 25 = USD 76.95; fine-tuning 6.00 + 24.00 + 400 = USD 430; 5.6 times. Value of 150 hours saved: USD 1,043 (150 × 6.95 = 1,042.50, rounded); RAG pays back in 4.1 to 10 months and fine-tuning in 9.8 to 24 months. With 0 changes per month, fine-tuning is cheaper per month.
Before tuning a model
Before asking for fine-tuning
Example pilot: RAG first, fine-tuning only if needed
- Week 1-2RAG and test questionsCurrent documents, real questions with their correct answer.KPI: question set approved by the business area
- Week 3-4Measure where it failsFix documents and search; adjust instructions.KPI: % knowledge failures vs. form failures
- Week 5-6DecideScale with RAG, or prepare fine-tuning if the form still fails.KPI: % correct and cost per query
More on how it works in what is RAG.