Pricing
You only pay for what you use on Replicate, billed by the second. When you don’t run anything, it scales to zero and you don’t pay a thing.
Hardware | Price | GPU | CPU | GPU RAM | RAM |
---|---|---|---|---|---|
CPU
cpu |
$0.000100/sec
$0.36/hr |
- | 4x | - | 8GB |
Nvidia A100 (80GB) GPU
gpu-a100-large |
$0.001400/sec
$5.04/hr |
1x | 10x | 80GB | 144GB |
2x Nvidia A100 (80GB) GPU
gpu-a100-large-2x |
$0.002800/sec
$10.08/hr |
2x | 20x | 160GB | 288GB |
4x Nvidia A100 (80GB) GPU
gpu-a100-large-4x |
$0.005600/sec
$20.16/hr |
4x | 40x | 320GB | 576GB |
8x Nvidia A100 (80GB) GPU
gpu-a100-large-8x |
$0.011200/sec
$40.32/hr |
8x | 80x | 640GB | 960GB |
Nvidia L40S GPU
gpu-l40s |
$0.000975/sec
$3.51/hr |
1x | 10x | 48GB | 65GB |
2x Nvidia L40S GPU
gpu-l40s-2x |
$0.001950/sec
$7.02/hr |
2x | 20x | 96GB | 144GB |
4x Nvidia L40S GPU
gpu-l40s-4x |
$0.003900/sec
$14.04/hr |
4x | 40x | 192GB | 288GB |
8x Nvidia L40S GPU
gpu-l40s-8x |
$0.007800/sec
$28.08/hr |
8x | 80x | 384GB | 576GB |
Nvidia T4 GPU
gpu-t4 |
$0.000225/sec
$0.81/hr |
1x | 4x | 16GB | 16GB |
Additional hardware | |||||
Nvidia H100 GPU
gpu-h100 |
$0.001525/sec
$5.49/hr |
Flux fine-tunes run on H100s; additional H100 capacity is reserved for committed spend contracts. | |||
2x Nvidia H100 GPU
gpu-h100-2x |
$0.003050/sec
$10.98/hr |
Flux fine-tunes run on H100s; additional H100 capacity is reserved for committed spend contracts. | |||
4x Nvidia H100 GPU
gpu-h100-4x |
$0.006100/sec
$21.96/hr |
Flux fine-tunes run on H100s; additional H100 capacity is reserved for committed spend contracts. | |||
8x Nvidia H100 GPU
gpu-h100-8x |
$0.012200/sec
$43.92/hr |
Flux fine-tunes run on H100s; additional H100 capacity is reserved for committed spend contracts. |
If you’re new to Replicate you can try featured models for free, but eventually you’ll need to enter a credit card.
Public models
Thousands of open-source machine learning models have been contributed by our community and more are added every day. When running or training one of these models, you only pay for time it takes to process your request.
Each model runs on different hardware and takes a different amount of time to run. You’ll find estimates for how much they cost under "Run time and cost" on the model’s page. For example, for stability-ai/sdxl:
This model costs approximately $0.0036 to run on Replicate, or 277 runs per $1, but this varies depending on your inputs.
This model runs on Nvidia L40S GPU hardware. Predictions typically complete within 4 seconds.
Image models
Audio models
Replicate hosts some audio models that are either priced per audio file, or per second of audio generated by the model.
Model | Output |
---|---|
minimax/music-01 | $0.035 / audio-file |
Model | Output |
---|---|
playht/play-dialog | $0.001 / second of audio |
Video models
Replicate hosts some video models that are either priced per video, or per second of video generated by the model.
Model | Output |
---|---|
luma/ray | $0.450 / video |
minimax/video-01 | $0.500 / video |
minimax/video-01-director | $0.500 / video |
minimax/video-01-live | $0.500 / video |
Model | Output |
---|---|
google/veo-2 | $0.500 / second of video |
haiper-ai/haiper-video-2 | $0.050 / second of video |
kwaivgi/kling-v1.6-pro | $0.098 / second of video |
kwaivgi/kling-v1.6-standard | $0.056 / second of video |
luma/ray-2-540p | $0.100 / second of video |
luma/ray-2-720p | $0.180 / second of video |
wavespeedai/hunyuan-video-fast | $0.200 / second of video |
wavespeedai/step-video | $0.250 / second of video |
wavespeedai/wan-2.1-i2v-480p | $0.150 / second of video |
wavespeedai/wan-2.1-i2v-720p | $0.500 / second of video |
wavespeedai/wan-2.1-t2v-480p | $0.120 / second of video |
wavespeedai/wan-2.1-t2v-720p | $0.450 / second of video |
Language models
Replicate hosts some language models that are priced per token.
Model | Input | Output |
---|---|---|
anthropic/claude-3.5-haiku | $1.000 / 1M tokens | $5.000 / 1M tokens |
anthropic/claude-3.5-sonnet | $3.750 / 1M tokens | $18.750 / 1M tokens |
anthropic/claude-3.7-sonnet | $3.000 / 1M tokens | $15.000 / 1M tokens |
deepseek-ai/deepseek-r1 | $10.000 / 1M tokens | $10.000 / 1M tokens |
ibm-granite/granite-20b-code-instruct-8k | $0.100 / 1M tokens | $0.500 / 1M tokens |
ibm-granite/granite-3.0-2b-instruct | $0.030 / 1M tokens | $0.250 / 1M tokens |
ibm-granite/granite-3.0-8b-instruct | $0.050 / 1M tokens | $0.250 / 1M tokens |
ibm-granite/granite-3.1-2b-instruct | $0.030 / 1M tokens | $0.250 / 1M tokens |
ibm-granite/granite-3.1-8b-instruct | $0.030 / 1M tokens | $0.250 / 1M tokens |
ibm-granite/granite-3.2-8b-instruct | $0.030 / 1M tokens | $0.250 / 1M tokens |
ibm-granite/granite-8b-code-instruct-128k | $0.050 / 1M tokens | $0.250 / 1M tokens |
meta/llama-2-13b | $0.100 / 1M tokens | $0.500 / 1M tokens |
meta/llama-2-13b-chat | $0.100 / 1M tokens | $0.500 / 1M tokens |
meta/llama-2-70b | $0.650 / 1M tokens | $2.750 / 1M tokens |
meta/llama-2-70b-chat | $0.650 / 1M tokens | $2.750 / 1M tokens |
meta/llama-2-7b | $0.050 / 1M tokens | $0.250 / 1M tokens |
meta/llama-2-7b-chat | $0.050 / 1M tokens | $0.250 / 1M tokens |
meta/meta-llama-3.1-405b-instruct | $9.500 / 1M tokens | $9.500 / 1M tokens |
meta/meta-llama-3-70b | $0.650 / 1M tokens | $2.750 / 1M tokens |
meta/meta-llama-3-70b-instruct | $0.650 / 1M tokens | $2.750 / 1M tokens |
meta/meta-llama-3-8b | $0.050 / 1M tokens | $0.250 / 1M tokens |
meta/meta-llama-3-8b-instruct | $0.050 / 1M tokens | $0.250 / 1M tokens |
mistralai/mistral-7b-instruct-v0.2 | $0.050 / 1M tokens | $0.250 / 1M tokens |
mistralai/mistral-7b-v0.1 | $0.050 / 1M tokens | $0.250 / 1M tokens |
A language model processes text by breaking it into tokens, or pieces of words. Replicate uses the Llama tokenizer to calculate the number of tokens in text inputs and outputs once it's finished.
Private models
You aren’t limited to the public models on Replicate: you can deploy your own custom models using Cog, our open-source tool for packaging machine learning models.
Unlike public models, most private models (with the exception of fast booting models) run on dedicated hardware so you don't have to share a queue with anyone else. This means you pay for all the time instances of the model are online: the time they spend setting up; the time they spend idle, waiting for requests; and the time they spend active, processing your requests. If you get a ton of traffic, we automatically scale up and down to handle the demand.
For fast booting models you'll only be billed for the time the model is active and processing your requests, so you won't pay for idle time like with other private models. Fast booting versions of models are labeled as such in the model's version list.
As with public models, if you would like more control over how a private model is run, you can use a deployments.
Learn more
For a deeper dive, check out how billing works on Replicate.
Enterprise & volume discounts
If you need more support or have complex requirements, we can offer:
- Dedicated account manager
- Priority support
- Higher GPU limits
- Performance SLAs
- Help with onboard, custom models, and optimizations
- Single sign-on
We've also got volume discounts for large amounts of spend. Email us at sales@replicate.com to learn more.