Fine-tuning vs RAG: which one does your task need, now that one door is closing (2026)
Data last updated: 2026-11-10. Pricing may have changed — verify with the vendor.
AITop is an independent guide. Prices and features change fast — always check the vendor's page.
New to the jargon? AITop's AI glossary explains it in plain English.
Related: what fine-tuning actually changes · what RAG is in buyer's terms · what embeddings cost · why the same prompt costs a tenth the second time
Fine-tuning retrains a model's weights on your examples; RAG (retrieval-augmented generation) fetches documents into the prompt at request time and leaves the model untouched. One changes the machine, the other changes what the machine reads. In 2026 that choice partly made itself: OpenAI announced it is winding down its self-serve fine-tuning platform — closed to new users, with the last new training jobs accepted January 6, 2027 per its own deprecations page — and Anthropic's official glossary states the Claude API does not currently offer fine-tuning at all. Google's Vertex AI still offers supervised tuning for a list of Gemini models. The comparison below is per the vendors' own published pages, checked November 2026.
RAG adds facts; fine-tuning changes behavior. If your model lacks knowledge, RAG is the answer — and in 2026 it is also the only half that every major vendor still sells intact. If your gap is format, tone or compliance, fine-tuning still works, but only where the door is still open: Google Vertex or open-weight models you train yourself.
| Fine-tuningchanges the weights | RAGchanges the context | |
|---|---|---|
| Pricing | OpenAI listed training at $25 per 1M tokens (gpt-4.1), $5 (gpt-4.1-mini), $3 (gpt-4o-mini) — and tuned inference costs more than base: fine-tuned gpt-4.1 runs $3.00 in / $12.00 out per 1M vs $2.00 / $8.00 untuned (+50%) | No training bill. A one-time embedding index (OpenAI lists $0.02–$0.13 per 1M tokens) plus ordinary per-request input tokens for the retrieved text — discounted when caching applies |
| Model | Bakes behavior into weights: format, tone, terminology, refusal patterns. Does not add facts — the knowledge cutoff is untouched (per OpenAI's own optimization guide and our fine-tuning entry) | Changes nothing about the model. Knowledge updates the moment your index updates — no retraining, no vendor schedule |
| Context window | Frees context: trained behavior needs no few-shot examples, so prompts shrink — which is how a +50% token premium can still net out cheaper at volume | Consumes context: every request re-pays input tokens for whatever the retriever fetched. Long documents at high volume is where the bill lives |
| Chinese support | Whatever your tuning examples demonstrate — a curated set per language is on you | Whatever your documents and retriever handle — multilingual by default if your index is |
| API | A dedicated model id (OpenAI's ft: prefix; Google Vertex returns a tuned model). In 2026, access is gated: OpenAI closed to new users with the last new jobs January 6, 2027; Anthropic offers none; Google Vertex tunes listed Gemini models (3.5 Flash / Flash-Lite, 3.1 Flash-Lite, 2.5 family) | No special endpoint — retrieval runs in front of any model you already call, on every vendor |
| Strengths |
|
|
| Weaknesses |
|
|
| Score | 7.8 | 9.1 |
What the landing pages won't tell you
The 2026 headline is not a price change — it is an exit. OpenAI announced the wind-down in May 2026: organizations that never fine-tuned were locked out May 7, organizations without recent tuned-model inference lost access July 2, and every remaining customer loses job creation January 6, 2027. Inference on existing fine-tunes continues until the base model is deprecated — and those dates are printed: October 23, 2026 for o4-mini and gpt-4.1-nano, January 14, 2027 for gpt-4o and gpt-4.1.
RAG is not free — it is unbundled. Embeddings are a one-time line item; retrieved tokens are a forever line item. The per-request retrieval bill is exactly what a fine-tuned model would have baked into weights, which is the honest way to price the comparison at your own query volume.
They are not rivals for the same job. RAG answers "what are the facts"; fine-tuning answers "how do you behave". OpenAI's own optimization guide runs evals → prompt engineering → fine-tuning as a loop, and the vendors' docs converge on one rule: fix the prompt, add retrieval for missing knowledge, and only fine-tune when a measured behavior gap survives both.
The verdict
In November 2026, default to RAG plus a better prompt, and let an evaluation set tell you when that isn't enough. Fine-tune only when a measured behavior gap — format, tone, classification discipline — survives good prompting and retrieval, and only on a platform still accepting jobs: Google Vertex tunes Gemini 3.5 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite and the 2.5 family; open-weight models you train yourself never had a door to close. If you already fine-tuned on OpenAI, note the dates: last new jobs January 6, 2027, and existing fine-tunes follow their base models — gpt-4o and gpt-4.1 included — out the door on January 14, 2027.
Straight answers
Can I still fine-tune OpenAI models in 2026?
Only if you already qualify. Per OpenAI's deprecations page, organizations that had never fine-tuned were shut out May 7, 2026; those without fine-tuned-model inference in the past 60 days lost access July 2, 2026; and active existing customers can create no new jobs after January 6, 2027. Fine-tuned models keep serving inference until their base model is deprecated — gpt-4o and gpt-4.1 base models shut down January 14, 2027.
Does Claude support fine-tuning?
No. Anthropic's official glossary states: "The Claude API does not currently offer fine-tuning, but ask your Anthropic contact if you are interested in exploring this option." Claude's behavior is adapted at request time — system prompts, examples in the prompt, retrieved documents, prompt caching.
Can I combine fine-tuning and RAG?
Technically yes — a fine-tuned model still accepts retrieved documents in its prompt. Practically, you then maintain both a training set and a retrieval index, and you pay both bills. Test the combination against RAG alone on a fixed evaluation set before adopting it; the vendors' own guidance is to reach for each tool only when the cheaper one measurably fails.