Published prices, open models
Together AI
The full rate card is public: per-token SFT and DPO prices by model, priced as dataset tokens × epochs, with small per-job minimums.
Guide · Fine-tuning services
Best Fine-Tuning Services (2026)
Fine-tuning is where annotation buyers end up: the model has to behave like your domain, not like the open internet. The market rearranged itself in 2025–2026 — two famous names were acquired and the largest one is closing — so this page starts from who can actually take your job today, at what published price.
| Provider | Model | What you can tune | Pricing | Best for |
|---|---|---|---|---|
| Together AI | Self-serve platform | Open models (Llama, Qwen, DeepSeek, gpt-oss) — SFT, DPO, LoRA and full fine-tuning | Published: SFT $0.34–$15 per 1M training tokens by model; DPO at 2.5× SFT; per-job minimums from $4 | Teams fine-tuning open models with no infrastructure and no sales call |
| Fireworks AI | Self-serve platform | Open models — fine-tuning API plus dedicated training on its GPU fleet | Published: training per token or per GPU-hour; H100 at $8/hour; zero data retention by default | Production teams that want training and fast serving from one provider |
| Microsoft Foundry (Azure OpenAI) | Cloud platform | OpenAI models (gpt-4.1 family, gpt-4o, o4-mini) and open models — SFT, DPO, reinforcement fine-tuning | Published: gpt-4.1 training $25 per 1M tokens, 4.1-mini $5; plus $1.70/hour hosting per deployment | Enterprises that need OpenAI models tuned inside their Azure boundary |
| Google Vertex AI | Cloud platform | Gemini, Gemma, Llama, Qwen — supervised, preference and reinforcement tuning | Published: $1.50–$25 per 1M training tokens by model; tuned endpoints bill at base-model inference prices | Google Cloud teams tuning Gemini itself |
| Amazon Bedrock | Cloud platform | Amazon Nova and open models — fine-tuning, continued pre-training, reinforcement fine-tuning | Published per model on AWS pricing pages | AWS-native teams that need customization under existing AWS terms |
| OpenAI (direct) | Closing | Existing customers only — no new fine-tuning users | Published while it lasts: gpt-4.1 training $25 per 1M tokens, 4.1-mini $5 | Nobody new — the platform is being wound down; migrate to Azure |
Training + serving, one bill
Fireworks AI
Fine-tune on its training API, serve on the same platform; zero data retention by default and a published compliance stack (SOC 2 Type II, ISO 27001, HIPAA BAA).
The data side of the job
Snorkel AI
When the platform is chosen and the dataset is the problem: expert data development — demonstrations, evaluations, preference data — as a service.
The market just changed shape
Three exits explain the table above. OpenAI is winding down its own fine-tuning platform — its pricing page says so in plain words — and pushing the workload to Azure. Predibase, the LoRA specialist whose LoRAX server made multi-adapter serving cheap, was acquired by Rubrik in 2025; its site now redirects to Rubrik and the self-serve platform is no longer sold under its own name. OpenPipe, the distillation-and-fine-tuning favorite, was acquired by CoreWeave in 2025 and its functionality migrated there. What replaced them is a split market: the three clouds tune their own models inside their own compliance boundaries, and open-model platforms (Together, Fireworks and their peers) compete on published per-token prices. Any "best fine-tuning platforms" list still leading with the departed names is describing a market that no longer exists.
What fine-tuning costs, in shape
Every provider on this page bills training the same way: training-dataset tokens × epochs, priced per million. That makes the training run easy to estimate and, for most teams, surprisingly small — tens to hundreds of dollars at open-model prices for a focused dataset. The costs that actually govern a fine-tuning program sit around the run:
- Hosting: a tuned model that must stay deployed bills by the hour. Azure\u2019s $1.70/hour per deployment is roughly $1,241 a month, forever, per fine-tune — the line item that quietly outruns every training run.
- Serving premium: some clouds bill tuned endpoints above base-model inference; Vertex bills tuned Gemini endpoints at base price, which is a genuine differentiator.
- The dataset: demonstrations and preference pairs written or graded by people who know the domain. Priced like expert annotation, because it is — see the cost guide.
- Iteration: the first tune is a draft. Eval sets, failure review and retraining are the program; budget for the loop, not the run.
The data decides
A fine-tuning platform will faithfully teach a model whatever is in your file — including your inconsistencies. The teams that get results treat the dataset as the product: demonstrations written to a rubric by people qualified in the domain, preference pairs graded against the same rubric, and an eval set held out before training starts. That is the human feedback market, and it is the same supply chain as the annotation companies this publication covers — one tier of expertise up. Expert networks are the direct route to demonstration writers: Micro1, for example, certifies experts through an AI-led interview process and matches them to post-training projects, with per-role pay ranges published on its job board. Companies can register to post projects: register your company with Micro1 →
Disclosure: this is a referral link. If your company registers with Micro1 through it, we may earn a referral commission at no cost to you. It does not change where Micro1 ranks — see our methodology.
Fine-tuning is not the only lever
Before commissioning a tune, run the honest comparison: half the "the model doesn\u2019t know our domain" problems are retrieval problems. Our RAG vs fine-tuning comparison works through which problems belong in the weights and which belong in a vector database — and why enterprise agents usually end with both. Regulated buyers should also read the enterprise fine-tuning guide: the compliance questions (retention, residency, weight ownership) matter more than the training price.
Frequently asked questions
What are the best fine-tuning services?
For self-serve fine-tuning of open models, Together AI and Fireworks AI publish real per-token training prices and run the job without a sales process. For fine-tuning OpenAI models, Microsoft Foundry (Azure) is now the route — OpenAI is winding down its own platform. Google Vertex AI tunes Gemini inside Google Cloud, and Amazon Bedrock customizes Nova and open models inside AWS. Which is "best" depends on which model you need tuned and which cloud boundary it has to stay inside.
How much does LLM fine-tuning cost?
Training is billed almost everywhere as training-dataset tokens multiplied by epochs, priced per million tokens. Published list prices run from $0.34 per million (Together AI, Llama 3.1 8B supervised fine-tuning) through $1.50–$5 (Vertex AI Gemini Flash tiers; Azure gpt-4.1-mini) up to $25 per million (gpt-4.1 on Azure and OpenAI; Gemini 2.5 Pro on Vertex). A worked example: a 10-million-token dataset trained for 3 epochs is 30 million training tokens — about $10 at Together’s 8B list price, $150 at a $5 rate, $750 at a $25 rate. The training run is usually the small bill; hosting the tuned model (Azure charges $1.70/hour per deployment, roughly $1,241 a month) and producing the training data are the larger ones.
Can I still fine-tune on OpenAI?
Not as a new customer. OpenAI’s own pricing page states it is winding down the fine-tuning platform: closed to new users, existing users can create jobs for a limited window, and fine-tuned models keep serving until their base models are deprecated. The same OpenAI models remain tunable through Microsoft Foundry on Azure.
What data does fine-tuning need?
Supervised fine-tuning learns from demonstrations: prompt-and-ideal-response pairs, typically hundreds to tens of thousands for a focused behavior. Preference tuning (DPO and its successors) learns from ranked pairs of outputs. The compute is the cheap part — producing demonstrations and preference data that are correct, consistent and rights-clean is the actual project, and it is the same supply chain as data annotation. See our guide to human feedback training services for how that data is bought.
Should I fine-tune, or use RAG?
RAG changes what the model can look up; fine-tuning changes how the model behaves. Facts that change belong in retrieval; consistent format, tone, tool-use and domain reasoning belong in the weights. Most production systems end up doing both — our RAG vs fine-tuning comparison works through the decision, including the enterprise-agent case.
Sources
Prices read from each provider’s own published pricing in November 2026; Azure figures are read from Microsoft’s published retail price list.
- Together AI pricing — fine-tuning per 1M training tokens (SFT and DPO tables)
- Fireworks AI pricing — training rates, GPU rates, data-retention and compliance posture
- OpenAI API pricing — fine-tuning wind-down notice and remaining training prices
- Vertex AI pricing — model tuning per 1M training tokens (Google Cloud)
- Azure OpenAI fine-tuning prices read from Microsoft’s retail price list — ecorp IT, November 2026
- Where OpenAI fine-tunes move — Foundry, Vertex, Bedrock and open-model platforms (Developers Digest, October 2026)
Facts on this page were drawn from the public sources above, checked November 2026. If you spot an error, tell us at the corrections address in the footer — we fix sourced errors and note the change.
Last updated: November 8, 2026