Fine-tuning vendor scorecard

Eight LLM fine-tuning vendors, scored on what their own pages publish about price, models, evaluation, data handling and deployment. Together AI leads on 7 of 10; data handling is unstated for all eight.

Independent: no company named here owns, sponsors or has endorsed this site. We have not tested any vendor: the method is based on published information only.

Last verified 7 November 2026 against each vendor's own pages

The scorecard

Sorted by score. Vendor names link to their notes.

Scored from each vendor's own published pages, read 7 November 2026. Not tested.
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: According to OpenAI's own API pricing page, new users can no longer sign up for fine-tuning. Teams already on it may keep creating training jobs "for the coming months"; their tuned models remain callable until each base model is deprecated. A new project cannot start there, so it carries no score.

What does the scorecard tell a buyer?

Price transparency is solved at the platform end. Together AI, Google Vertex AI, Fireworks AI, AWS Bedrock all publish training prices with units. The trouble is that the units differ: per million tokens, per 1,000 tokens, per training hour and per GPU-hour all appear.

Serving terms are published where training is. The same vendors that publish training prices also state how a tuned model is served and on what basis, from Fireworks AI's "no extra charge" (you pay for GPU time) to Azure OpenAI's hourly hosting line.

Evaluation is rarely part of the published product. Only Together AI (billed validation runs) and Prime Intellect (evaluation environments) mention it on the pages we read. Plan to bring your own test set.

Data handling is the open question. None of the pricing or fine-tuning pages we read states retention, training-use or deletion terms. Ask for the data processing agreement before you upload anything.

Why each square is filled or empty

The evidence line for every criterion, taken from the vendor's own page.

Vendor	Published pricing	Model choice	Evaluation support	Data handling	Deployment
Together AI	Per-token SFT and DPO prices for each model, with a minimum charge per job.	Llama, Qwen, Gemma, Mixtral, gpt-oss, DeepSeek, Kimi and GLM models listed.	Validation tokens from evaluation runs are part of the billed token count.	No retention or training-use terms on the pricing pages we read.	Serving uses dedicated endpoints, charged by the minute for the hardware chosen (docs).
Google Vertex AI	Training prices for Gemini and open models, per 1,000 or per million tokens.	Gemini 3.5, 3.1 and 2.5, plus Gemma 3, Llama 3.x and 4 Scout, and Qwen 3.	Not covered on the pricing page we read.	Not covered on the pricing page we read.	Tuned Gemini 3+ priced as the base model in the global region, 1.5x elsewhere.
Fireworks AI	LoRA and full SFT and DPO prices by base-model size, plus GPU-hour rates.	Four size tiers from up to 16B to over 300B parameters; LoRA and full tuning.	Not covered on the pricing page we read.	Not covered on the pricing page we read.	No extra charge to serve a tuned model; deployments billed per GPU-second.
AWS Bedrock	Per-hour reinforcement fine-tuning and per-token fine-tuning prices published.	gpt-oss-20b, Qwen3 32B and Llama models, with Nova, Titan and Cohere tables too.	Not in the part of the pricing page we captured.	Not in the part of the pricing page we captured.	On-demand per-token or provisioned per-hour serving, plus monthly model storage.
Azure OpenAI	Fine-tuning price lines are listed, but figures load in the browser and were not captured.	OpenAI models only: o4-mini, GPT-4.1, 4.1-mini, 4.1-nano, GPT-4o and 4o-mini.	Not covered on the pricing page we read.	Not covered on the pricing page we read.	Tuned models carry an hourly hosting charge alongside training.
Anyscale	Custom quote; no fine-tuning prices published.	Docs cover LoRA or full-weight runs with SFT, DPO or RLVR; base models not recorded.	Not covered in the docs we read.	Not covered in the docs we read.	Not covered in the docs we read.
Prime Intellect	No prices on the pages we read.	Not covered on the pages we read.	Hosts a public hub where RL and eval environments are built and shared.	Not covered on the pages we read.	Not covered on the pages we read.
OpenPipe	Its own product pages were not read in this pass.	Its own product pages were not read in this pass.	Its own product pages were not read in this pass.	Its own product pages were not read in this pass.	Its own product pages were not read in this pass.

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.

Sources and dates read (10)

Pricing, Together AI (viewed 7 Nov 2026)
Fine-tuning pricing, Together AI docs (viewed 7 Nov 2026)
Generative AI on Vertex AI pricing, Google Cloud (viewed 7 Nov 2026)
Pricing, Fireworks AI (viewed 7 Nov 2026)
Bedrock pricing, AWS (only part of the page was recorded) (viewed 7 Nov 2026)
Azure OpenAI Service pricing, Microsoft (figures did not load) (viewed 7 Nov 2026)
LLM fine-tuning, Anyscale docs (viewed 7 Nov 2026)
Environments Hub announcement, Prime Intellect blog (viewed 7 Nov 2026)
OpenPipe acquisition news, CoreWeave (viewed 7 Nov 2026)
API pricing, OpenAI (viewed 7 Nov 2026)

Questions about the scorecard

What does an empty square mean?

That nothing on the vendor pages we checked (all on 7 November 2026) stated it. It does not mean the vendor lacks the feature. Where we did not read a vendor's own product pages at all (OpenPipe), the vendor notes say so.

Why is OpenAI not on the scorecard?

New sign-ups for OpenAI's own fine-tuning have stopped, according to its API pricing page, which describes the platform as winding down. A scorecard for teams starting a project should not rank an option they cannot sign up for.

Is a higher score a better vendor?

A higher score means more of what a buyer needs to compare is published. It is a measure of transparency, not of model quality, which only your own evaluation can show.

How often is the scorecard checked?

Every vendor page is re-read on each check, and changes are dated in the terms changelog.