| Base model | $ / 1M tokens |
|---|---|
| Less than 4B parameters | $0.10 |
| 4B - 16B parameters | $0.20 |
| More than 16B parameters | $0.90 |
| MoE 0B - 56B parameters (e.g. Mixtral 8x7B) | $0.50 |
| MoE 56.1B - 176B parameters (e.g. DBRX, Mixtral 8x22B) | $1.20 |
| DeepSeek V3 family | $0.56 input, $1.68 output |
| DeepSeek R1 0528 | $1.35 input, $5.4 output |
| GLM-4.5, GLM-4.6 | $0.55 input, $2.19 output |
| Meta Llama 3.1 405B | $3.00 |
| Meta Llama 4 Maverick (Basic) | $0.22 input, $0.88 output |
| Meta Llama 4 Scout (Basic) | $0.15 input, $0.60 output |
| Qwen3 235B Family, GLM-4.5 Air | $0.22 input, $0.88 output |
| Qwen3 30B, Qwen Coder Flash | $0.15 input, $0.60 output |
| Kimi K2 Instruct, Kimi K2 Thinking | $0.60 input, $2.50 output |
| Qwen3 Coder 480B | $0.45 input, $1.80 output |
| OpenAI gpt-oss-120b | $0.15 input, $0.60 output |
| OpenAI gpt-oss-20b | $0.07 input, $0.30 output |
Discounts for prompt caching are available for enterprise deployments. Contact us to learn more.
Batch inference is priced at 50% of our serverless pricing for both input and output tokens. Learn more here.
Pay per second of audio input
| Model | $ / audio minute (billed per second) |
|---|---|
| Whisper-v3-large | $0.0015 |
| Whisper-v3-large-turbo | $0.0009 |
| Streaming ASR v1 | $0.0032 |
| Streaming ASR v2 | $0.0035 |
| Image model name | $ / step | Approx $ / image |
|---|---|---|
| All Non-Flux Models (SDXL, Playground, etc) | $0.00013 per step ($0.0039 per 30 step image) | $0.0002 per step ($0.006 per 30 step image) |
| FLUX.1 [dev] | $0.0005 per step ($0.014 per 28 step image) | N/A on serverless |
| FLUX.1 [schnell] | $0.00035 per step ($0.0014 per 4 step image) | N/A on serverless |
| FLUX.1 Kontext Pro | $0.04 per image | N/A |
| FLUX.1 Kontext Max | $0.08 per image | N/A |
All models besides the Flux Kontext models are charged by the number of inference steps (denoising iterations). The Flux Kontext models are charged a flat rate per generated image.
| Base model parameter count | $ / 1M input tokens |
|---|---|
| up to 150M | $0.008 |
| 150M - 350M | $0.016 |
| Qwen3 8B | $0.1 |
Priced per 1M training tokens
| Base Model | Supervised Fine Tuning | Direct Preference Optimization |
|---|---|---|
| Models up to 16B parameters | $0.50 | $1.00 |
| Models 16.1B - 80B | $3.00 | $6.00 |
| Models 80B - 300B (e.g. Qwen3-235B, gpt-oss-120B) | $6.00 | $12.00 |
| Models >300B (e.g. DeepSeek V3, Kimi K2) | $10.00 | $20.00 |
SFT and DPO prices are shown in $ per 1M training tokens. Training tokens can be estimated with number of tokens in training dataset * number of epochs. Estimation should be multiplied by the average number conversation turns /2 for tuning with intermediate thinking traces.
Fine-tuning with images (VLM supervised fine-tuning) is also billed per 1M tokens. See this FAQ on calculating image tokens.
Reinforcement fine tuning jobs are priced per GPU hour (billed per second), at the same price as Fireworks on-demand deployment. Please see the section below for details on RFT pricing.
| GPU Type | $ / hour (billed per second) |
|---|---|
| A100 80 GB GPU | $2.90 |
| H100 80 GB GPU | $4.00 |
| H200 141 GB GPU | $6.00 |
| B200 180 GB GPU | $9.00 |
For estimates of per-token prices, see this blog. Results vary by use case, but we often observe improvements like ~250% higher throughput and 50% faster speed on Fireworks compared to open source inference engines.
