Together AI
Inference, Tuning and GPU Clusters
API
Fireworks AI
Serverless inference, managed fine-tuning with published per-token prices, and on-demand GPUs. Note that GPU rates rise from 1 September 2026.
What Is Fireworks AI?
Fireworks AI runs open models in production. Serverless inference for the common path, managed fine-tuning when a general model is not specific enough, and dedicated GPUs when you need hardware that belongs to you for the hour.
Its distinguishing habit is publishing the numbers other vendors keep behind a form. The fine-tuning price list is a grid — model size against training method — with a figure in every cell, which means you can estimate a training run from your dataset before speaking to anybody. In a market where training costs are usually quoted on request, that is a genuine difference in kind.
One thing to know before you plan around any figure here: the company has published a price increase for on-demand GPUs taking effect on 1 September 2026. Rates rise across the range, in some cases by 30%. Anything you cost out on the current numbers needs checking against the new ones.
How It Works
Serverless First
The default path is per-token inference across open models, with $1 in free credits to begin. That is a small amount deliberately — enough to confirm the API works and not enough to defer a decision about payment.
Embeddings are priced separately and cheaply, banded by model size: $0.008 per million input tokens up to 150M parameters, $0.016 from 150M to 350M, and $0.1 for Qwen3 8B. For retrieval systems, where embedding volume dwarfs generation volume, those bands are often the number that decides the architecture.
Fine-Tuning With a Price List
Managed training is billed per million training tokens, and the grid is worth reading properly because the numbers scale steeply. LoRA supervised tuning costs $0.50 up to 16B parameters, $3.00 from 16.1B to 80B, $6.00 from 80B to 300B and $10.00 above that. DPO doubles each figure, and full-parameter tuning doubles it again.
Read that as a twentyfold spread between the cheapest and most expensive supervised LoRA, driven entirely by base model size. The practical advice this implies is unglamorous: try the small model first. A great many fine-tuning projects that were budgeted against a 70B base would have worked on an 8B one, and the difference is $0.50 against $3.00 per million tokens.
On-Demand GPUs, and the September Increase
When serverless is not enough you can take GPUs by the hour. Through 31 August 2026 an H100 or H200 is $7.00, a B200 $10.00, a B300 $12.00 and a GB300 $18.00. From 1 September those become $8.00, $13.00, $15.00 and $20.00 respectively.
The B200 increase is the sharpest at 30%, and it is the one most likely to matter, since B200s are the current workhorse for serious inference. Deployments restricted to a specific region carry a 1,5× premium on top, which is easy to overlook when data residency is added late in a project as a compliance requirement rather than a budget item.
What It Costs
Two price lists matter: what training costs per million tokens, and what a GPU costs per hour. The second is the one changing.
| What | Price | Notes |
|---|---|---|
| Free credits | $1 | Evaluation only |
| Embeddings up to 150M params | $0.008 | Per million input tokens |
| Embeddings 150M–350M params | $0.016 | Per million input tokens |
| Embeddings, Qwen3 8B | $0.1 | Per million input tokens |
| LoRA SFT up to 16B | $0.50 | Per million training tokens |
| LoRA SFT 16.1B–80B | $3.00 | Per million training tokens |
| LoRA SFT 80B–300B | $6.00 | Per million training tokens |
| LoRA SFT above 300B | $10.00 | Per million training tokens |
| Full-parameter SFT up to 16B | $1.00 | Double the LoRA rate |
| DPO | Double the SFT rate | At every size band |
| H100 / H200 GPU | $7.00 → $8.00 | Per hour, rising 1 September 2026 |
| B200 GPU | $10.00 → $13.00 | Per hour, rising 1 September 2026 |
| B300 GPU | $12.00 → $15.00 | Per hour, rising 1 September 2026 |
| GB300 GPU | $18.00 → $20.00 | Per hour, rising 1 September 2026 |
| Region-restricted deployment | 1.5× | Premium on GPU rates |
| Enterprise | Contact sales | Custom terms |
The doubling structure on training is easy to remember and easy to misjudge. Full-parameter DPO on a model above 300B is $40.00 per million tokens against $10.00 for the LoRA equivalent — a four-fold difference for a choice that is often made casually at the start of a project.
Who Gets the Most From It
Teams Budgeting a Fine-Tune
Published per-token training prices across every size band mean you can produce a credible estimate from your dataset size in an afternoon. That is unusual and it is the strongest reason to look here.
Retrieval-Heavy Applications
Cheap banded embedding pricing suits systems that index far more text than they generate. At $0.008 per million tokens for smaller models, embedding a large corpus stops being the line item that kills the project.
Production Workloads Needing Dedicated Hardware
Hourly GPUs with a clear upgrade path from serverless, and a range that reaches current-generation Blackwell parts. Model the September rates rather than the current ones if the deployment is not immediate.
Anyone With Data Residency Requirements
Region-restricted deployment is available, which many inference vendors do not offer at all. Price it at 1,5× from the beginning rather than discovering the multiplier during a compliance review.
What to Watch Out For
- GPU prices rise on 1 September 2026, by up to 30%. Any comparison written before that date — including cost models you built yourself — is now out of date.
- One dollar of free credit is a handshake, not a trial. Plan to pay before you have finished evaluating.
- Training costs scale twentyfold with base model size. Choosing the base casually is the single most expensive decision in a fine-tuning project.
- Region restriction costs 1,5×. Compliance requirements arriving late will move your budget more than most people expect.
- Open models only. As with Together and Groq, frontier commercial reasoning is not part of the catalogue.
- Hourly GPUs bill whether or not you use them. The move from serverless is worth calculating rather than assuming.
Frequently Asked Questions
Are Fireworks prices going up?
On-demand GPU rates yes, from 1 September 2026. H100 and H200 go from $7.00 to $8.00 an hour, B200 from $10.00 to $13.00, B300 from $12.00 to $15.00 and GB300 from $18.00 to $20.00. Serverless and fine-tuning prices are quoted separately from this change.
How much does fine-tuning actually cost?
Multiply your training tokens by the rate for your base model size and method. A LoRA supervised run on a 13B model with 10 million training tokens is roughly $5. The same run on a 200B base is around $60.
Is there a free tier?
There are $1 in free credits, which is enough to confirm your integration works. It is not a free tier in the sense that Groq offers one.
How does it compare to Together AI?
They overlap heavily. Together reaches further down into raw GPU clusters with long-term reserved pricing; Fireworks publishes a clearer fine-tuning grid. If training economics are your main question, start here; if you expect to rent clusters, start there.
Can I deploy in a specific region?
Yes, at a 1,5× premium on the GPU rate. That is a supported option rather than an enterprise-only conversation, which is worth knowing if residency is a requirement and your volume is modest.
What is the difference between LoRA and full-parameter tuning?
LoRA adjusts a small set of added weights and costs half as much; full-parameter tuning updates the whole model. LoRA is the sensible default and is usually sufficient for teaching a model a domain, a format or a tone.
The Bottom Line
Fireworks AI is the inference platform that will tell you what things cost before you commit. The fine-tuning grid alone justifies a look, because it turns a question normally answered by a sales call into arithmetic you can do yourself.
Do that arithmetic against the post-September GPU rates, not the current ones, and try the smallest base model your task will tolerate. The platform is straightforward and the pricing is honest; the mistakes here are the ones you make by picking a bigger model than you needed.