LLM Fine-Tuning Services
Most model problems are solved by retrieval or a better prompt, and the rest by fine-tuning. Find your situation in the table, then see what the work costs at published rates.
| Your situation | Usually the right approach | Typical cost basis |
|---|---|---|
| Answers must come from your own documents, and those documents change | Retrieval (RAG) firstNo training run | Embedding and vector storage, then per-token inference on a base model. Longer prompts cost more per request. |
| The model knows enough, but the format, tone or structure is wrong | Prompt with examples, then supervised fine-tuning (SFT)If prompting still drifts | Training charged on tokens processed: every token in the file, once for each epoch.1 |
| One narrow task (classify, route, extract) at high daily volume | Fine-tune a small open modelUsually with LoRA | Published rates from $0.34 to $0.50 per million training tokens for models up to 16B, then serving billed per GPU-second or per minute.2 |
| Reviewers can rank two answers and you want the better kind every time | Preference tuning (DPO)After an SFT pass | Priced per training token at about two to two and a half times the SFT rate.2 |
| Success can be checked by a program (tests pass, the task completes) | Reinforcement fine-tuningNeeds a grader you trust | Billed per training hour on some platforms: $80.00 an hour on one major cloud.3 |
| A few hundred examples at most, and no test set yet | Prompting, while you build the evaluation setTraining comes later | Base-model inference only. |
- Google Vertex AI pricing (training tokens are dataset tokens times epochs) and Together AI pricing.
- Fireworks AI pricing and Together AI pricing.
- Amazon Bedrock pricing, reinforcement fine-tuning of gpt-oss-20b and Qwen3 32B.
Each page read 7 Nov 2026. List prices, not quotes. A provider named in these sources may pay us a fee for an introduction; that has no bearing on the figures.
Prompt, retrieve, then tune: in that order
Each step costs more to build and less to run than the one before. Move down only when the step above has been measured and found short.
Prompting
Start here, always
Clear instructions and a few worked examples sent with each request. It is the fastest way to learn what the task actually needs, and the baseline every other option has to beat on the same test set.
Cost: per-token inference only. It grows with prompt length and traffic.
Retrieval (RAG)
When facts are missing or stale
Search your documents at answer time and put the best passages in the prompt. Facts stay current by updating documents, not models, and every answer can point to its source.
Cost: embedding, a vector index and longer prompts. Most of the effort goes into cleaning documents and tuning the search.
Fine-tuning
When behavior is the problem
Train the model (or a small adapter on top of it) on examples of the output you want. It fixes format, style and narrow judgment, and lets a small model do one job a large model was doing at higher cost per request.
Cost: a training job billed per token or hour, an evaluation set, and serving the tuned model.
Retrieval and fine-tuning are not rivals: many production systems tune a model to read retrieved passages well. The fine-tuning vs RAG guide compares them question by question.
What does a training run cost at published rates?
Per million training tokens, for a small and a mid-size open model on three platforms. A 2 million token dataset trained for 3 epochs is 6 million training tokens: about $3.00 at $0.50 per million.
| Provider | About 8B | What that covers | About 70B | What that covers |
|---|---|---|---|---|
| Fireworks AI | $0.50 | LoRA SFT, models up to 16B | $3.00 | LoRA SFT, 16.1B to 80B |
| Together AI | $0.34 | SFT, Llama 3.1 8B ($4.00 minimum per job) | $2.03 | SFT, Llama 3.3 70B |
| Google Vertex AI | $0.67 | Supervised tuning, Llama 3.1 8B | $6.72 | Supervised tuning, Llama 3.3 70B |
Disclosure: a provider in this table can pay us a referral fee when one of our introductions turns into paid work. Rows, their order and every figure are set without regard to that.
OpenAI is missing on purpose: its price list (7 Nov 2026) says its fine-tuning is being retired and is shut to new customers. Full tables for Fireworks AI, Together AI, Vertex AI, AWS Bedrock and the rest are on the prices page, and the project scoper shows which rate applies to your model size.
Who does fine-tuning work?
Platforms where your engineers run the job, and specialists for one stage of it. Alphabetical; what each publishes is on the providers page.
| Provider | Type | Published pricing |
|---|---|---|
| Anyscale | Compute platform (Ray) | Not published for fine-tuning |
| AWS Bedrock | Cloud model platform | Per 1M training tokens (Llama 2) or $80.00 per training hour (reinforcement fine-tuning) |
| Azure OpenAI | Cloud model platform | Token-based training (hourly for o4-mini); amounts not captured |
| Fireworks AI | Inference and training platform | $0.50 to $10.00 per 1M training tokens for LoRA SFT, by model size |
| Google Vertex AI | Cloud model platform | Gemini per 1,000 training tokens; open models per 1M |
| Together AI | Inference and fine-tuning cloud | From $0.34 per 1M tokens (SFT, Llama 3.1 8B); minimum $4.00 to $60.00 per job |
| OpenPipe (CoreWeave) | Reinforcement learning for agents | Not published |
| Prime Intellect | Open RL training and environments | Not published |
| SuperAnnotate | Data platform for training sets | Three tiers (Starter, Pro, Enterprise), amounts unlisted |
Disclosure: some firms in this list can pay us a referral fee when one of our introductions turns into paid work. Inclusion, order and wording are decided without regard to that.
What does a fine-tuning project need before it starts?
Three things decide whether a tuned model ships. None of them is the training job.
- Examples of the output you want
Input and ideal-output pairs that look like real traffic. Narrow tasks can move on a few hundred clean pairs; broad tasks need thousands. A person should have checked them.
- An evaluation set
Real inputs held out from training, with agreed correct answers and a scoring method. Without it, nobody can say the tuned model is better than a good prompt.
- A hosting plan
Shared serving, dedicated GPUs or your own account. Dedicated capacity is billed while idle, which can outweigh training many times over.
Project checklist
- One sentence of scope. The task the model does, for whom, and what a wrong answer costs.
- A baseline. Prompting and, where facts matter, retrieval, scored on the evaluation set.
- A clean training set. Checked examples, personal details removed, rights to use them confirmed.
- Base model and license. The exact version, and whether its license allows your commercial use.
- Method. LoRA or full fine-tuning; SFT, DPO or reinforcement fine-tuning.
- Where it runs. Training and serving locations agreed with your security team.
- Ownership in writing. Who holds the tuned weights, whether they can be exported, and whether the provider keeps your data.
- A running cost at your volume. Serving priced at expected traffic, idle hours included.
- A retraining plan. What triggers the next run and who re-checks the evaluation.
All prices here are list figures with the date they were copied; treat none as a quote for your work. Providers are described from what they publish, without testing or audits by us.
Questions about LLM fine-tuning services
What do LLM fine-tuning services include?
At the light end, a platform that runs the training job on your data and serves the tuned model. At the full end, a firm that helps define the task, builds the training and evaluation sets, runs the tuning, tests the result and hands over a model you can deploy. Most projects buy the platform and do some of the data work in-house.
Should we fine-tune or use RAG?
Use retrieval when the model needs facts from your documents, and fine-tuning when the facts are fine but the behavior is wrong. Many production systems use both. Our fine-tuning vs RAG guide has the side-by-side table.
How much does fine-tuning cost?
The training job is often a few dollars to a few hundred at published rates; the work around it costs more. Data preparation, an evaluation set and hosting the tuned model are where the budget goes. The price tables list each provider's rates with units.
How many examples do we need?
Enough to show the task clearly and to hold back a test set. Narrow tasks can respond to a few hundred clean examples; broad ones need thousands. Quality matters more than count. The data volume post explains how to size it.
Can we still fine-tune GPT models on OpenAI?
Not as a new customer. A notice on OpenAI's price list (seen 7 Nov 2026) says its fine-tuning is being retired. Azure OpenAI still lists tuning for GPT-4.1 and GPT-4o models, and Vertex AI tunes Gemini.
Do we have to send you our data?
Never. The outline describes the task in words and is all this site handles. Any data moves later, between your team and the chosen provider, under the contract you sign with them.
Sources behind this section
- Fireworks AI, price list (figures copied 7 Nov 2026)
- Together AI, price list (figures copied 7 Nov 2026)
- Vertex AI (Google Cloud), generative AI price list (figures copied 7 Nov 2026)
- Amazon Bedrock, price list (copied 7 Nov 2026, page only partly captured)
- OpenAI, API price list (figures copied 7 Nov 2026)
- Azure OpenAI, price list (figures copied 7 Nov 2026)
Get a quote
Only three things are needed: the company's ZIP code, where to send our answer, and a short sketch of the project in taps. The answer names providers whose published offers line up with it.
Keep data out of this form: no samples, records, weights or keys. Describing the task in words gives us enough to work with.
This site is an introduction service, and a provider it introduces you to may pay it a fee. Your details stay with us until you approve a named provider.
Received. Thanks for the outline.
- Next we set it against the published offers of fine-tuning providers: their models, methods and deployment options.
- An email follows naming the ones that fit, with a reason for each. Your details go nowhere until you approve one.
- From there, that provider deals with your team directly and prices the job. This site plays no part in the contract.