Analyst note, November 2026
Eight vendors, five criteria, published pages only
Together AI and Fireworks AI publish the most usable fine-tuning terms of the eight vendors we scored: per-token training prices, broad open-model menus and a stated price basis for serving the result.
Google Vertex AI and AWS Bedrock match them on price transparency and keep the tuned model inside a cloud account you may already run.
None of the eight states data retention terms on the pages we read, and OpenAI has stopped taking new fine-tuning customers on its own platform.
Get a quote$0.50
Two points for a criterion published in full, one for partly, none where the pages we read say nothing. Last verified 7 November 2026.
| Vendor | Published pricing | Model choice | Evaluation support | Data handling | Deployment | Score |
|---|---|---|---|---|---|---|
| Together AIOpen-model cloud, API | Published in full | Published in full | Partly published | Not found on the pages we read | Published in full | 7 / 10 |
| Google Vertex AICloud platform | Published in full | Published in full | Not found on the pages we read | Not found on the pages we read | Published in full | 6 / 10 |
| Fireworks AIOpen-model cloud, API | Published in full | Published in full | Not found on the pages we read | Not found on the pages we read | Published in full | 6 / 10 |
| AWS BedrockCloud platform | Published in full | Published in full | Not found on the pages we read | Not found on the pages we read | Published in full | 6 / 10 |
| Azure OpenAICloud platform | Partly published | Partly published | Not found on the pages we read | Not found on the pages we read | Published in full | 4 / 10 |
| AnyscaleCompute platform (Ray) | Not found on the pages we read | Partly published | Not found on the pages we read | Not found on the pages we read | Not found on the pages we read | 1 / 10 |
| Prime IntellectRL training and environments | Not found on the pages we read | Not found on the pages we read | Partly published | Not found on the pages we read | Not found on the pages we read | 1 / 10 |
| OpenPipeAgent RL training (CoreWeave) | Not found on the pages we read | Not found on the pages we read | Not found on the pages we read | Not found on the pages we read | Not found on the pages we read | 0 / 10 |
Published in full (2 points)Partly published (1)Not found on the pages we read (0)
Vendors on this list can owe this site a fee if an introduction it makes ends in a signed contract. Squares, order and wording are set without regard to that.
Not scored: OpenAI, whose pricing page reads "OpenAI is winding down the fine-tuning platform." New users cannot sign up. Open the full scorecard for the reason behind every square.
One line per vendor from its own pages. The vendor notes carry the detail and sources.
Open-model cloud, API, 7 / 10
The widest open-model menu on our list, from sub-1B Qwen3.5 models up to Qwen3.5 397B A17B, DeepSeek and GLM-5.x. Its pricing page shows SFT at $0.34 per million tokens for Llama 3.1 8B and $2.03 for Llama 3.3 70B, with DPO at $0.84 and $5.08.
Cloud platform, 6 / 10
Google's Gemini models and open models (Gemma, Llama, Qwen) are tuned from the same price page. Methods vary by model: supervised tuning across the list, reinforcement learning fine-tuning on Gemini 3.5 Flash, and preference tuning on Gemini 2.5 Flash and Flash Lite.
Open-model cloud, API, 6 / 10
The clearest price sheet of the eight: managed training is priced by base-model size, with LoRA SFT from $0.50 per million training tokens (up to 16B) to $10.00 (over 300B), and full SFT at double the LoRA rate in each tier.
Cloud platform, 6 / 10
Reinforcement fine-tuning of OpenAI gpt-oss-20b or Qwen3 32B is priced at $80.00 per training hour, and the tuned models are then served on demand per token. Older Llama 2 fine-tuning is priced per million training tokens ($1.49 for 13B, $7.99 for 70B) and served on provisioned throughput at $23.50 per model unit per hour without commitment.
Cloud platform, 4 / 10
The Azure OpenAI price list we read covers OpenAI models only: reinforcement fine-tuning for o4-mini (per training hour plus hosting per hour) and supervised tuning for the GPT-4.1 and GPT-4o families (per million training tokens plus hosting per hour).
Compute platform (Ray), 1 / 10
A Ray-based compute platform rather than a fine-tuning API. According to its docs, post-training runs on Ray Train, SkyRL or LLaMA-Factory, and can be supervised, DPO or reinforcement learning with verifiable rewards (RLVR), as LoRA adapters or full-weight updates.
RL training and environments, 1 / 10
Prime Intellect launched its Environments Hub on 27 August 2025: an open platform for building and sharing reinforcement learning and evaluation environments. It sits closer to the reinforcement fine-tuning end of the market than to supervised tuning.
Agent RL training (CoreWeave), 0 / 10
CoreWeave announced on 3 September 2025 that it was acquiring OpenPipe, which it described as a leader in reinforcement learning.
From published information only. For each vendor we read its own pricing page and, where one exists, its fine-tuning documentation, then fill a square only for what those pages state. We have not run jobs on any platform or tested a tuned model.
Vendors quote faster and more honestly when these are settled first. None of them requires sending data.
The task in one sentence
What the tuned model must do better than the base model, stated so an outsider could judge it.
A baseline and a target
Today's score with prompting or retrieval, and the score that would justify tuning.
Base model family
Closed API, open-weight, or must run on your hardware. This alone rules out half the list.
Examples on hand
Roughly how many reviewed input and output pairs exist, and who can label more.
Where the model must run
The vendor's endpoint, your cloud account or on premises, plus expected monthly traffic.
Data rules
Retention limits, regions and whether training data may be used by the vendor for anything else.
Review a vendor must pass
A SOC 2 report, a data processing agreement or your own questionnaire.
Who owns the result
Tuned weights or adapters, the evaluation set and the grader, if one is used.
Have most of these? We can tell you which vendors' published terms line up with them.
Get a quoteTraining prices copied from each vendor's own page on 7 November 2026. Units differ, so read the last column before comparing. Published prices, not quotes.
| Vendor | Model or tier | Method | Price | Unit |
|---|---|---|---|---|
| Fireworks AI | Base model up to 16B | LoRA SFT | $0.50 | per 1M training tokens |
| Fireworks AI | Base model up to 16B | Full SFT | $1.00 | per 1M training tokens |
| Fireworks AI | Base model 16.1B to 80B | LoRA SFT | $3.00 | per 1M training tokens |
| Fireworks AI | Dedicated H100 80GB | Training compute | $8.00 | per GPU-hour |
| Together AI | Llama 3.1 8B | SFT | $0.34 | per 1M tokens |
| Together AI | Llama 3.1 8B | DPO | $0.84 | per 1M tokens |
| Together AI | Llama 3.3 70B | SFT | $2.03 | per 1M tokens |
| Together AI | gpt-oss-120B | SFT | $2.50 | per 1M tokens |
| Google Vertex AI | Llama 3.1 8B | SFT | $0.67 | per 1M training tokens |
| Google Vertex AI | Llama 3.3 70B | SFT | $6.72 | per 1M training tokens |
| Google Vertex AI | Gemini 3.5 Flash | SFT | $0.01 | per 1,000 training tokens |
| Google Vertex AI | Gemini 2.5 Flash Lite | SFT / preference | $0.0015 | per 1,000 training tokens |
| AWS Bedrock | Qwen3 32B | RFT (reinforcement) | $80.00 | per training hour |
| AWS Bedrock | Llama 2 70B, pretrained | Standard tuning | $7.99 | per 1M training tokens |
| OpenAIClosed to new users | gpt-4.1-mini | SFT | $5.00 | per 1M training tokens |
Independent: no company named here owns, sponsors or has endorsed this site. Vendors on this list can owe this site a fee if an introduction it makes ends in a signed contract. Squares, order and wording are set without regard to that.
It depends on the base model and where the model must run. On published information alone, Together AI scores highest of the eight (7 of 10), with Google Vertex AI, Fireworks AI and AWS Bedrock next on 6. The full scorecard shows why each square is filled.
Not on OpenAI's own platform. Its API pricing page (viewed 7 November 2026) describes fine-tuning as winding down and no longer open to new sign-ups. Azure OpenAI still lists fine-tuning for OpenAI models.
Published training prices start at $0.34 per million tokens (Together AI, Llama 3.1 8B, SFT) and $0.50 (Fireworks AI, LoRA SFT up to 16B). Billed tokens are your dataset multiplied by epochs, and serving the tuned model is a separate, recurring cost.
None of the pricing and documentation pages we read state how long training data is kept, whether it is used to train other models, or how it is deleted. Those terms usually sit in data processing agreements: ask each vendor for them in writing.
No. We read each vendor's own published pages and score what they state. We do not run training jobs, benchmark models or audit security. The scoring method lists exactly what counts.
Use a platform if you already have clean examples and a test set. Hire help when the hard part is the data, the evaluation or the deployment. Our post on consultancy or provider tools sets out the signals.
Start with a ZIP code, add an email, then tap through what you want tuned. We read it, check which vendors' published terms fit, and write back with names before anyone else sees your brief.
This is an introduction service. The vendor you are introduced to may pay it a fee if you sign work together; the team asking is never charged.