About Replicate
Replicate runs machine learning models behind a cloud API, so the user never provisions or manages a GPU. Models are packaged with Cog, an open source tool that wraps a model in a container with a generated API server. A deployment scales with demand and scales down to zero when it is idle.
The HTTP API sits at https://api.replicate.com/v1 and authenticates with an API token sent in the Authorization header using the Bearer scheme. Endpoints cover predictions, models, model versions, deployments, trainings, hardware and account, and a running prediction or training can be stopped through a dedicated cancel endpoint. Webhooks give asynchronous completion notifications and streaming output is supported. Client libraries are published for Python and Node.js alongside the plain HTTP API. Creating a prediction is rate limited to 600 requests per minute and every other endpoint to 3,000, with an HTTP 429 returned past either limit.
Fine tuning is done by calling the trainings API with a base model and a training dataset, which produces a new deployable model version. Model output files are served from the replicate.delivery domain and its subdomains, and prediction inputs, outputs and logs created through the API are deleted after one hour by default.
Billing is per second of compute at a rate set by the hardware chosen. Standard CPU compute is $0.0001 per second, or $0.36 an hour, an Nvidia T4 is $0.000225 or $0.81 an hour, an Nvidia L40S is $0.000975 or $3.51 an hour, and a single Nvidia H100 is $0.001525 or $5.49 an hour. An Nvidia A100 with 80GB of memory costs $5.04 an hour, and an eight way H100 configuration costs $0.0122 per second, which works out at $43.92 an hour. Some models are priced per unit of output instead: Flux 1.1 Pro at $0.04 per generated image, Flux Schnell at $3.00 per thousand images, and Wan 2.1 image to video at 720p at $0.25 per second of output. Private models are billed for their entire running time, including boot and idle, while fast booting fine tunes are billed only for active processing.

Key features
- Runs machine learning models behind a cloud API without provisioning or managing GPUs.
- Bills compute per second at a rate set by the hardware selected.
- Prices hardware from $0.09 per hour on small CPU to $43.92 per hour on eight Nvidia H100s.
- Prices some models per unit of output, such as $0.04 per generated image.
- Prices video models per second of output, such as $0.25 per second at 720p on Wan 2.1.
- Scales a model deployment with demand and down to zero when it is idle.
- Fine tunes a base model through the trainings API, producing a new deployable version.
- Packages models with Cog, an open source tool that wraps a model in a container with an API server.
- Exposes endpoints for predictions, models, model versions, deployments, trainings, hardware and account.
- Cancels a running prediction or training through a dedicated endpoint.
- Notifies on completion through webhooks and streams output as it is generated.
- Provides Python and Node.js client libraries alongside a plain HTTP API.
- Authenticates with an API token sent as a bearer token in the Authorization header.
- Limits create prediction calls to 600 requests per minute and other endpoints to 3,000.
- Deletes prediction inputs, outputs and logs one hour after creation by default.
What people use it for
Key facts
| Category | AI |
|---|---|
| Pricing | Paid |
| Free tier | No |
| Upvotes | 12 |
| Listed | 22 August 2026 |
Is this your product?
Replicate alternatives
- Open source
- Freemium
- Open source
- Freemium
More in AI
OpenAI's assistant, with browsing, images and code execution built in.
The model hub. Weights, datasets and demos for most of open-source ML.
Speech synthesis and voice cloning that survives being listened to closely.
Image generation with a house style nothing else quite matches.
An answer engine that cites its sources instead of listing ten blue links.
Video generation and editing built for people who already edit video.
Anthropic's assistant, strongest on long documents and careful reasoning.