Replicate
Replicate is a cloud platform that runs, fine-tunes and deploys AI models through a single API. Its catalogue spans more than 50,000 models for images, video, speech, music and text, billed per second of compute rather than by subscription.
What is Replicate?
Replicate is a hosted inference platform. Its terms define the service as a SaaS platform that lets customers run models in the cloud, either through the Replicate API or through the website, for a usage-based fee. The pitch on the homepage is deliberately narrow: run and fine-tune models, deploy custom models, all with one line of code. The documentation restates the same idea from the developer's side, promising that you can run machine learning models with a cloud API without understanding the intricacies of machine learning or managing your own infrastructure.
The catalogue is the centre of gravity. The enterprise page counts more than 50,000 models, mixing official releases from Google, OpenAI, Anthropic, ByteDance, Black Forest Labs, Alibaba, xAI, MiniMax, ElevenLabs, Meta, Stability AI, Ideogram, Recraft, Krea and Pruna with thousands of community contributions. The Explore page organises them by task: text to image, text to video, image to video, text to speech, speech to text, OCR, super resolution, image restoration, background removal, video editing, vision, text classification, speaker diarisation, face swap and NSFW detection, among others.
Three layers of use sit on top of that catalogue. The simplest is calling a published model, which the official Python, Node.js, Go and Swift clients reduce to a single run call. The second is fine-tuning, where you supply a dataset and a trigger word and get back a model addressable like any other. The third is bringing your own code: Cog, Replicate's open-source packaging tool, turns a model into a container with a generated API server and deploys it on the company's cluster.
Operationally, the platform scales up automatically under load and down to zero when idle, which is what makes the per-second billing meaningful. Predictions carry explicit statuses and time out after thirty minutes; output files are served from replicate.delivery. Registration goes through GitHub only. Replicate, LLC is based in San Francisco; Cloudflare announced its acquisition of the company in November 2025, with the brand kept in place.
What it does
- Run any of more than 50,000 hosted AI models from a single API call
- Compare candidate models side by side in the Playground before writing code
- Fine-tune an image model on your own dataset and call the result like any other model
- Package and deploy a custom model with Cog, Replicate's open-source model packaging tool
- Create a deployment to pin hardware, minimum and maximum instances behind a stable endpoint
- Receive results asynchronously through webhooks or stream outputs as they are produced
- Drive the whole API from a coding assistant through the MCP server or the published agent skills
When to use Replicate / When not to
A quick filter to help you decide if Replicate is the right fit.
When to use Replicate
- Software engineers who want to ship an AI feature without provisioning GPUs, wrestling with CUDA or hosting model weights themselves
- Product teams prototyping with several models at once, comparing candidates in the Playground before committing to one in code
- Startups with spiky or unpredictable traffic, since public models scale down to zero and idle time is not billed
- Creative and media teams generating or editing images, video, speech and music programmatically at volume
- Machine learning practitioners who want to package a custom model with Cog and expose it as a production API without building the serving layer
When not to use Replicate
- Non-technical users looking for a ready-made interface: everything runs through the API, a client library or a bare testing view
- Anyone who cannot sign up with a GitHub account, since that is the only registration path offered
- Organisations that require European data residency, as every subprocessor Replicate publishes is located in the United States
- Buyers who need a documented GDPR position, a named Article 27 representative or a published security certification before signing
- Teams that need a mobile app or a fixed monthly budget: there is no iOS or Android app, and billing is metered per second of compute
How to use Replicate
A typical end-to-end flow, from setup to results.
- Sign in with a GitHub account, which is the only registration route the platform offers
- Browse the Explore page or open the Playground to compare candidate models on your own prompts
- Create an API token from the account settings
- Install an official client (pip install replicate, npm install replicate) or call the HTTP API directly
- Run a first prediction, for example replicate.run("black-forest-labs/flux-schnell", input={"prompt": "an astronaut riding a horse"})
- For long-running models, poll the prediction endpoint or register a webhook to collect results
- Add prepaid credit and configure auto-reload before any sustained usage, since credit must be bought upfront
- To fine-tune, call replicate.trainings.create with a destination model, a base version, a zip of images and a trigger word
- To ship your own model, write a cog.yaml describing the environment and a predict.py exposing setup and predict, then push it
- Create a deployment to fix the hardware and the minimum and maximum instance counts behind a stable endpoint
Pros & Cons
Pros
- One API token unlocks tens of thousands of models from competing providers, and switching between them is a one-line change
- No GPU provisioning, CUDA management or model weight hosting to deal with
- Genuinely usage-based billing: public models charge only for active time, setup and idle are free, and failed runs are not billed
- Extensive documentation, exportable as markdown, with an llms.txt index and an OpenAPI schema
- Cog is open source, so the packaging work done for Replicate is not locked to Replicate
- API predictions have their inputs, outputs and logs purged automatically after one hour
- A published subprocessor list, a public changelog and a public status page make the operational posture easy to audit
Cons
- Registration is restricted to GitHub accounts, with no email sign-up path
- No documented way to exclude customer data from model training, and the terms never state whether Replicate trains its own models on it
- No GDPR compliance claim, no Article 27 representative and no published certification such as SOC 2 or ISO 27001
- All published subprocessors sit in the United States, so there is no European data residency option
- Prepaid credit must be bought upfront, expires after one year and is not refundable
- Costs are hard to forecast: private models and deployments are billed for setup and idle time as well as active time
- No public support email address; support runs through an authenticated page, Discord or X
Pricing & Plans
There is no permanent free plan. A selection of models can be tried at no charge, but the billing documentation states that customers are asked to set up billing after a short while. Beyond that, Replicate is strictly pay-as-you-go: compute is metered per second and the rate depends on the hardware, from 0.000025 USD per second (0.09 USD per hour) on a small CPU instance to 0.012200 USD per second (43.92 USD per hour) on eight Nvidia H100 or H200 GPUs. Some models are priced per unit of output instead, for example 0.025 USD per image on flux-dev or 3.00 USD per million input tokens on Claude 3.7 Sonnet. Credit is purchased upfront, remains valid for one year and is not refundable.
- billed only for the time the model is actively processing your request
- with setup and idle time free
- billed for setup
- idle and active time on dedicated hardware
- billed only for active processing time
- whether the model is public or private
- credit bought upfront and drawn down as you go
- valid one year
- non-refundable
- with optional auto-reload from a 5 USD threshold and a 15 USD minimum top-up
- usage from the previous month charged at the start of the next one
- quote-based
- adding a dedicated account manager
- priority support
- higher GPU limits
- performance SLAs
- onboarding and custom model assistance
- contractual indemnity and volume discounts
Data, GDPR & hosting
A consolidated view of how Replicate handles your data.
GDPR overview
There is no GDPR statement on this site. Searching the whole collected corpus, including the full documentation index, returns no compliance claim, no dedicated GDPR page, no Article 27 EU representative and no named data protection officer. The regulation appears once, inside the supplemental terms for Black Forest Labs models, purely as an obligation imposed on the customer. The privacy policy does grant access, deletion, correction and marketing opt-out rights, but it frames them without reference to European law and adds a notice specific to US states only. Governing law is Californian, the venue is San Francisco, and the seventeen published subprocessors are all located in the United States. European buyers should treat the GDPR position as undocumented and raise it directly with the vendor.
Who owns the data?
Section 5.1 of the terms is explicit: the customer owns all rights, title and interest, including intellectual property rights, in its Customer Data, defined as the inputs it submits and the outputs the models return. Replicate additionally assigns to the customer whatever rights it may hold in the outputs, commercial use included, subject to the third-party terms attached to the model that produced them. In exchange, section 5.2 grants Replicate a worldwide, non-exclusive, royalty-free licence limited to delivering the service. Replicate keeps ownership of the platform itself, of any feedback submitted, and of Resultant Data, which the terms define as aggregated and anonymised usage information.
Reuse rights
Outputs can be reused without asking Replicate for permission. Section 5.1 states that the customer may exploit them commercially, including by selling or publishing them, with one important qualification: the third-party terms of whichever model generated the output still apply, so the licence attached to a given model on the marketplace governs what can legitimately be done with its results. Publishing a model publicly is a different matter: under section 5.2 a Community Model carries a perpetual, irrevocable, worldwide licence for every other user of the service, and Replicate may make it private or remove it at its own discretion. Feedback sent to the company is assigned outright and cannot be reclaimed.
Data retention & training
Hosting summary
Replicate publishes a subprocessor list in its documentation and locates every one of the seventeen entries in the United States. AWS is named for cloud infrastructure and data hosting; GCP, CoreWeave and Fly.io provide additional compute; Crunchy Bridge runs Postgres, TigerData handles time-series storage and CloudAMQP the message queue. Cloudflare provides content delivery, Honeycomb tracing, Sentry error monitoring and Metabase analysis. Stripe processes payments, Metronome handles usage invoicing, Attio the CRM, Pylon customer support, Loops.so marketing email and LaunchDarkly feature flags. Output files are served from replicate.delivery and its subdomains. No European region is offered and no data residency option is advertised anywhere on the site. The domain resolves to a Cloudflare anycast address geolocated in the United States, which identifies the CDN edge rather than an origin server, so it should not be read as evidence about where data actually sits.
Where Replicate works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting Replicate.
- Prepaid credit expires after one year and is not refundable, so buying a large balance ahead of uncertain usage carries real risk
- Private models and deployments are billed for idle time as well as active time: an instance left running quietly keeps charging
- Predictions created through the web interface are kept indefinitely, unlike API predictions, which is easy to forget when testing with sensitive inputs
- Outputs remain subject to the third-party licence of the model that produced them, so commercial use needs a per-model check rather than a blanket assumption
- No documented training opt-out and no stated position on whether the vendor trains on customer inputs: sensitive material deserves caution
- The terms impose mandatory individual arbitration under Californian law, with waivers of class action and jury trial
- Convenience can erode judgement: swapping models in one line encourages shipping generated output without evaluating quality, bias or provenance
Setup & Integrations
Technical difficulty
Low for a developer, moderate for anyone shipping their own model. Running a published model takes a GitHub sign-in, an API token, a client library and a single call, with no machine learning knowledge required. Fine-tuning an image model is barely harder: a zip of images, a trigger word and a step count. Deploying custom code is the real step up, requiring a cog.yaml describing the GPU and Python environment, a predict.py exposing setup and predict, and some familiarity with containers. Long-running predictions also need polling or webhook handling. Step-by-step guides cover Node.js, Python, Google Colab, Next.js, SwiftUI and CI/CD.
Deployment
Integrations
Supported languages
Behind Replicate
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Do I need a particular kind of account to use Replicate?
How does Replicate charge for usage?
Is there a free plan?
How long does Replicate keep the data I send to a model?
Who owns the images, text or audio the models produce?
Are my inputs used to train AI models?
Where is my data hosted?
What is the minimum age to use the service?
Are there API rate limits?
Can I deploy my own model rather than using the catalogue?
Should you pick Replicate?
Replicate solves a narrow problem well: it removes the infrastructure layer between a developer and a working AI model. One token, one API call, and a catalogue of more than 50,000 models covering images, video, speech, music, text and vision. The per-second billing is real rather than cosmetic, since public models charge only for active processing, failed runs are not billed and idle capacity scales to zero. Cog, the open-source packaging tool, means the work of containerising a custom model is not wasted if you later leave the platform. Documentation is thorough, exportable and machine-readable, and the MCP server and agent skills make the whole API usable from a coding assistant.
The reservations are mostly contractual rather than technical. There is no GDPR statement anywhere on the site, no Article 27 representative, no published certification and no European hosting region: every one of the seventeen disclosed subprocessors is American. The terms grant a licence covering the training of customer derivative models but say nothing about whether Replicate trains its own models on customer inputs, and no opt-out is documented. Predictions created in the web interface are kept indefinitely, unlike API predictions which are purged after an hour. Prepaid credit expires after a year and is not refundable, and private models and deployments are billed for idle time, so an unattended instance can cost money without producing anything.
It is a strong fit for engineering teams that want breadth of models and elastic cost, and a poor one for organisations with a formal data protection checklist. The acquisition by Cloudflare, announced in November 2025 with the brand retained, is worth watching: it will shape where the platform runs and under which contracts.
- Choosing a selection results in a full page refresh.
- Opens in a new window.