Pricing snapshot
Fireworks AI uses usage-based pricing across three main paths: serverless inference, managed or serverless training, and on-demand GPU deployments. Serverless inference uses prepaid credits: add a payment method, purchase credits, and usage for input, output, and cached tokens is deducted from the balance. Auto Reload can replenish a low balance, while a monthly spend limit can cap usage.
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Workload | Published rate |
|---|---|
Embedding models up to 150M parameters | $0.008 per 1M input tokens |
Embedding models from 150M to 350M parameters | $0.016 per 1M input tokens |
Qwen3 8B embeddings | $0.10 per 1M input tokens |
Choose a billing path
Serverless inference: for token-metered access with no setup or cold starts. Serving paths include Standard, Priority, and Fast; Fireworks directs buyers to its pricing documentation for current popular-model prices.
Training: managed supervised and preference fine-tuning is billed per 1M training tokens. The Serverless Training API bills separately for prefilling, cached prefilling, sampling, and training tokens.
On-demand deployments: dedicated capacity is billed per GPU second, with no extra charge for start-up time. Use this path when dedicated deployments, higher rate limits, or scale-oriented GPU pricing fit the workload.
Published rate details
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| GPU type | Price per minute | Price per hour |
|---|---|---|
H100 80 GB GPU | $0.134 | $8.00 |
H200 141 GB GPU | $0.134 | $8.00 |
B200 180 GB GPU | $0.217 | $13.00 |
B300 288 GB GPU | $0.250 | $15.00 |
GB300 288 GB GPU | $0.334 | $20.00 |
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Base-model size | LoRA SFT | LoRA DPO | Full parameter SFT | Full parameter DPO |
|---|---|---|---|---|
Up to 16B parameters All prices are per 1M training tokens. | $0.50 | $1.00 | $1.00 | $2.00 |
16.1B to 80B parameters All prices are per 1M training tokens. | $3.00 | $6.00 | $6.00 | $12.00 |
80B to 300B parameters All prices are per 1M training tokens. | $6.00 | $12.00 | $12.00 | $24.00 |
Over 300B parameters All prices are per 1M training tokens. | $10.00 | $20.00 | $20.00 | $40.00 |
Cost drivers to check
Managed-training estimates start with training-dataset tokens multiplied by epochs. For tuning with intermediate thinking traces, Fireworks says to also multiply by the average number of conversation turns divided by two; reasoning-content fields can increase tuned-token totals.
For Serverless Training API jobs, prefill, cached prefill, sample, and train are separate chargeable operations. Checkpoint storage for serverless models is included during private preview.
A region-restricted on-demand deployment is priced at a 1.5× premium, so the deployment region remains part of the capacity estimate.
Fireworks invites buyers to contact sales for enterprise deployments, describing faster speeds, lower costs, and higher rate limits for that option.
Take
Choose Fireworks AI when token-metered serverless access, token-priced training, or dedicated GPU deployments match how you intend to run models. Before committing, map the workload to its operation—input, output, cached input, prefill, sample, training token, or GPU runtime—and confirm the selected serving path, model, training method, and region.