Replicate — run, fine-tune, and deploy AI models behind one API

Use when

You want one hosted API over thousands of community and official models—call a model by name from Python, Node, or HTTP, fine-tune it on your own data, or package custom code with Cog and let shared capacity scale to zero between requests.

Skip when

You need a predictable per-output price on a shortlist of media models, dedicated hardware pinned to a region you choose, or a click-through app instead of an API and a token.

Try instead

Compare a media-focused inference API when per-output pricing matters more than catalog breadth, or rent raw GPUs when you would rather run the serving stack yourself.

• fal.ai overview
• AutoDL overview

• ReplicateThis page

  Pricing

  Prepaid credit drawn down by usage: most public models bill per second of active hardware time, official models bill per output or input, and private deployments also bill setup and idle time

  Learning cost

  One HTTP endpoint or the Python and Node clients; packaging your own weights with Cog adds a container build step

  Output limits

  Broadest model spread, including niche research and community checkpoints; official models stay warm while community models can cold-boot

  Privacy

  Input parameters, outputs, output files, and logs from API predictions are removed about an hour after they run; predictions created in the web interface persist until you delete them

• fal.ai

  Pricing

  Prepaid credits for serverless model APIs, billed mostly per output, plus hourly GPU rates for compute

  Learning cost

  Call a model endpoint from REST or an SDK client; the curated catalog keeps initial setup shallow

  Output limits

  Catalog concentrates on production generative media—image, video, audio, and 3D—rather than long-tail research weights

  Privacy

  Inputs and outputs pass through fal infrastructure—review its retention and residency terms for your data class

• Hugging Face Inference Endpoints

  Pricing

  Paid plan required first, then per-minute billing while an endpoint is initializing or running; the hourly rate follows the instance type you select

  Learning cost

  Pick a Hub model, then configure instance type, scaling, and serving stack—the most ops-heavy of the three

  Output limits

  Bounded by whichever model you deploy and the GPU you attach, covering text, embedding, and diffusion workloads

  Privacy

  Endpoints run on dedicated instances you deploy, with the provider and region chosen at setup

1. Replicate — official site
2. Replicate — pricing
3. Replicate — prediction data retention
4. fal.ai — official site
5. Hugging Face — Inference Endpoints

Free tier

Yes

Pricing summary

Replicate is usage-billed rather than subscription-billed. New accounts add prepaid credit and usage draws it down, with overages settled at the start of the following month. Most public models charge only for the time they spend actively processing a request, while official models bill per output or per input and private deployments also charge for setup and idle time. A small set of models can be run free before billing is required, and prepaid credit has an expiry window. Confirm current rates and terms on the pricing page—this page does not list currency amounts.

Reviewed on 17 September 2026 · Replicate — pricing

• One API for any model

  Community and official models share the same prediction endpoint, so adding a model is a version string change rather than a new integration.

• Python, Node, and HTTP clients

  Maintained open-source clients wrap the REST API, which is the documented path for running models from scripts, notebooks, or backend services.

• Official models that stay warm

  Around a hundred maintained models run always-on with a stable input and output schema and metered per output rather than per second.

• Fine-tuning on your own data

  Train a tuned version of supported models on your own examples, then call it by name like any other model in the catalog.

• Custom models with Cog

  Package your own code and weights with Cog, an open-source container format, and Replicate generates the API server and scales it for you.

• Deployments for dedicated capacity

  Create a deployment when you need your own instances and request queue instead of the shared public pool, with control over hardware and scaling.

• Webhooks and streaming

  Send prediction events to an HTTPS endpoint, or consume server-sent events to render partial output while a longer job is still running.

1. 1

   Create an API token

   Sign in at replicate.com, open account settings, and generate a token; store it as REPLICATE_API_TOKEN in your environment.

2. 2

   Find the model identifier

   Browse the catalog and copy the owner/name identifier for the model, or pin a specific version when you need stable behavior across releases.

3. 3

   Call it from your code

   Install the Python or Node client and pass the model identifier plus an input object; file inputs accept a local path or an HTTPS URL.

4. 4

   Save output before it expires

   API predictions are cleaned up about an hour after they run, so copy files and metadata you want to keep or receive them through a webhook.

5. 5

   Fine-tune or package your own model

   Train on your own examples, or wrap bespoke weights with Cog and deploy them when the existing catalog does not cover the job.

6. 6

   Move to dedicated capacity when needed

   Add a deployment when shared queue latency or cold boots stop being acceptable for production traffic.

• Application engineers shipping AI features

  Integrate a model behind an HTTP call instead of provisioning and monitoring GPU servers.

• ML engineers and researchers

  Reach community checkpoints and niche research weights that a closed model API would never carry.

• Teams prototyping model pipelines

  Chain predictions together with webhooks and swap model versions without rewriting integration code.

• Indie builders and early startups

  Start on prepaid credit and scale on demand rather than reserving hardware you may not saturate.

API-first product

The primary interface is an HTTP API with Python and Node clients; the web interface is for exploring models and testing predictions.

Prediction data retention

Input parameters, outputs, output files, and logs from API predictions are removed after about an hour by default. Predictions created through the web interface are kept until you delete them.

Model licenses vary

Community models ship under their own licenses. Check the license listed on a model page before shipping it in a commercial product.

Public versus private infrastructure

Public models share capacity and scale down when idle; deployments and private models hold dedicated instances, which is why idle and setup time is billed for them.

Streaming and long jobs

Server-sent event streams are available when a version supports them and carry a request timeout, so long-running jobs are better tracked with webhooks or polling.

Hardware menu

Runs span CPU and GPUs from small inference cards up to multi-A100 configurations, and the per-second rate follows whichever hardware a model selects.