LLM fine-tuning prices by provider
Training a small open model costs cents to a few dollars per million training tokens on the platforms below, and larger models run from about $2 to $40 per million. Units differ by provider (per 1,000 tokens, per 1 million tokens, per training hour), so compare like with like. These are published list prices, not quotes.
Last verified , each line against the company's own pricing or product page.
We run this page on our own: none of the companies on it is a partner, owner or sponsor, and none has endorsed it. Every figure was copied from the provider's price list on 7 November 2026. When a page would not load in full, the gap is marked instead of filled.
Jump to: Fireworks AI, Together AI, Google Vertex AI, AWS Bedrock, OpenAI, Azure OpenAI, price changes log.
Fireworks AI: price per million training tokens, by model size
Managed training on open models, priced by the size of the base model. Full fine-tuning costs twice LoRA in every tier, and DPO twice SFT.
| Base model size | LoRA SFT | LoRA DPO | Full SFT | Full DPO |
|---|---|---|---|---|
| Up to 16B | $0.50 | $1.00 | $1.00 | $2.00 |
| 16.1B to 80B | $3.00 | $6.00 | $6.00 | $12.00 |
| 80B to 300B | $6.00 | $12.00 | $12.00 | $24.00 |
| Over 300B | $10.00 | $20.00 | $20.00 | $40.00 |
Hourly GPU rental for training is also listed: H100 80GB $8.00, H200 141GB $8.00, B200 180GB $13.00, B300 288GB $15.00, GB300 288GB $20.00. Serving a trained model carries no surcharge; the deployment is metered in GPU-seconds and the training job is billed on its own.
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.
Together AI: SFT and DPO price per million tokens
Together counts billable tokens as epochs times training tokens, plus the number of evaluation passes times validation tokens. Each job has a minimum charge.
| Model | SFT | DPO | Minimum per job |
|---|---|---|---|
| Llama 3.1 8B | $0.34 | $0.84 | $4.00 |
| Qwen3.5 0.8B / 2B / 4B / 9B | $0.34 | $0.84 | $4.00 |
| gpt-oss-20B | $0.40 | $1.00 | $4.00 |
| Qwen3.5 27B / 35B A3B, Qwen3.6 27B / 35B A3B | $1.05 | $2.62 | $4.00 |
| Gemma 4 26B A4B / 31B / 31B VLM | $1.05 | $2.62 | $4.00 |
| Mixtral 8x7B Instruct v0.1 | $1.05 | $2.62 | $4.00 |
| Llama 3.3 70B | $2.03 | $5.08 | $4.00 |
| Llama 4 Scout | $2.10 | $5.25 | $6.00 |
| gpt-oss-120B | $2.50 | $6.25 | $6.00 |
| Qwen3.5 122B A10B | $4.20 | $10.50 | $10.00 |
| Llama 4 Maverick | $5.60 | $14.00 | $16.00 |
| Qwen3.5 397B A17B | $5.60 | $14.00 | $22.00 |
| DeepSeek V4 Flash 0731 | $6.00 | $15.00 | $12.00 |
| DeepSeek-V3.1 | $7.00 | $17.50 | $20.00 |
| Kimi K2.6 / Kimi K2.7 Code | $15.00 | $37.50 | $60.00 |
| GLM-5.1 / GLM-5.2 | $40.00 | $100.00 | $60.00 |
Its documentation adds that a job which fails is refunded in full, and that a few models carry no minimum. To serve a tuned model, Together runs a dedicated endpoint charged by the minute at a rate that depends on the hardware chosen, on top of the training bill.
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.
Google Vertex AI: Gemini and open-model tuning prices
Gemini tuning is listed per thousand training tokens, open models per million. Beside each Gemini figure we add our own per-million conversion so it can be read against the other tables.
| Gemini model | Method | Published, per 1,000 tokens | Per 1M (our conversion) |
|---|---|---|---|
| Gemini 3.5 Flash Lite | Supervised (SFT) | $0.003 | $3.00 |
| Gemini 3.5 Flash | Supervised (SFT) | $0.01 | $10.00 |
| Gemini 3.5 Flash | Reinforcement (RFT) | $0.01 | $10.00 |
| Gemini 3.1 Flash Lite | Supervised (SFT) | $0.003 | $3.00 |
| Gemini 2.5 Pro | Supervised (SFT) | $0.025 | $25.00 |
| Gemini 2.5 Flash | Supervised, or preference | $0.005 | $5.00 |
| Gemini 2.5 Flash Lite | Supervised, or preference | $0.0015 | $1.50 |
| Open model (supervised tuning) | Per 1M training tokens |
|---|---|
| Llama 3.2 1B | $0.28 |
| Gemma 3 1B IT | $0.47 |
| Llama 3.2 3B | $0.61 |
| Llama 3.1 8B | $0.67 |
| Gemma 3 4B IT | $1.14 |
| Qwen 3 4B | $1.35 |
| Gemma 3 12B IT | $1.82 |
| Qwen 3 8B | $4.18 |
| Llama 4 Scout 17B 16E | $5.77 |
| Qwen 3 32B | $6.57 |
| Llama 3.3 70B | $6.72 |
| Gemma 3 27B IT | $6.83 |
| Qwen 3 14B | $8.46 |
Serving a tuned Gemini model: generation 3 onward costs the same as the untuned model when called through the global endpoint, and half as much again (1.5x) through a specific region. Tuned 2.5-and-earlier models simply inherit the base prediction rate.
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.
AWS Bedrock: customization prices we could read
Rates for US East (N. Virginia) and US West (Oregon). Only part of the price list loaded for us: anything for Llama 3.x, Nova, Titan or Cohere models, and distillation, is missing here and should be read on Amazon's page.
| Model | Method | Training | Serving the tuned model | Storage |
|---|---|---|---|---|
| OpenAI gpt-oss-20b | Reinforcement tuning (RFT) | $80.00 for each hour of training | $0.09 in, $0.39 out, per million tokens | $1.95 monthly, per model |
| Qwen3 32B | Reinforcement tuning (RFT) | $80.00 for each hour of training | $0.20 in, $0.78 out, per million tokens | $1.95 monthly, per model |
| Llama 2 Pretrained 13B | Supervised customization | $1.49 / million tokens trained | $23.50 an hour per model unit, provisioned without commitment | $1.95 monthly, per model |
| Llama 2 Pretrained 70B | Supervised customization | $7.99 / million tokens trained | $23.50 an hour per model unit, provisioned without commitment | $1.95 monthly, per model |
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: fine-tuning prices while the platform winds down
According to its own pricing page, OpenAI is retiring fine-tuning. New accounts can no longer start; current customers keep the ability to launch jobs for some months yet, and models already tuned keep working until the underlying base model is retired.
| Model | Training | Tuned input | Cached input | Tuned output | Queued input | Queued output |
|---|---|---|---|---|---|---|
| gpt-4.1-2025-04-14 | $25.00 | $3.00 | $0.75 | $12.00 | $1.50 | $6.00 |
| gpt-4.1-mini-2025-04-14 | $5.00 | $0.80 | $0.20 | $3.20 | $0.40 | $1.60 |
OpenAI also offers cheaper inference on a tuned model if the customer opts in to data sharing when starting the job.
Azure OpenAI: prices not published here
Microsoft's price list names tuning for o4-mini (reinforcement, charged by training hour, then hosting by the hour) and for GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, GPT-4o-2024-08-06 and GPT-4o-mini (charged per million training tokens, then hourly hosting). The amounts are drawn in by script after the page loads; we could not capture them and print none. Read them on the Azure OpenAI pricing page for your region.
Price changes log
Dated notes on changes to these providers' published tuning terms, newest at the top. A figure is only changed after its source page has been read again.
- : OpenAI added a notice that fine-tuning is being retired. No new customers; current ones may still launch jobs for some months, and tuned models run until their base model is retired.
- : Baseline: we read every price list on this page for the first time (Fireworks AI, Together AI, Vertex AI, Bedrock, OpenAI, Azure OpenAI).
Sources behind this section
- Fireworks AI, price list (figures copied 7 Nov 2026)
- Together AI, price list (figures copied 7 Nov 2026)
- Together AI, fine-tuning billing docs (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 (opened 7 Nov 2026, figures not captured)
Questions about fine-tuning prices
What does a single fine-tuning run cost?
Usually less than people expect. Four million training tokens on a small open model at Fireworks AI's adapter rate ($0.50 per million, models to 16B) comes to $2.00. Preparing the data, building the evaluation and paying for serving afterward almost always dwarf that.
What is a training token?
A token from your training file, counted again on every epoch. Vertex AI multiplies dataset tokens by epochs; Together AI also adds the validation tokens used in each evaluation pass. Two million tokens of data run for three epochs bills as six million.
Why is DPO priced higher than SFT?
DPO learns from a preferred and a rejected answer side by side, so each step processes more. Fireworks AI doubles its SFT price for DPO in each tier; on Together AI the DPO figure runs at roughly 2.5 times SFT.
Can we still fine-tune on OpenAI?
Only as an existing customer, and not for long. OpenAI's price list (7 Nov 2026) says fine-tuning is being retired and new customers cannot start. Plan a new project on another platform.
Do these prices include hosting the tuned model?
No. Each platform meters serving its own way: tokens for tuned Gemini on Vertex AI, GPU-seconds on Fireworks AI, minutes of a dedicated endpoint on Together AI, and model-unit hours for provisioned Llama 2 on Bedrock.
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.