Replicate review: Pay-per-use cloud API for running, fine-tuning and deploying ML models

What Replicate does

Lets developers run, fine-tune and deploy machine learning models for images, video, speech, music and text through a cloud API, billed by usage.

Replicate is a hosted service for developers who want to call machine learning models over an API rather than set up GPU servers. The homepage shows code samples in Node, Python and HTTP, and describes running a model with a single line of code.

Its catalog is made up of models published by a community and by vendors, covering image generation, speech, music, image restoration, video from images, captioning and large language models. Besides running existing models, you can fine-tune one with your own data, for example to produce images of a particular person, object or style. You can also bring your own model: Cog, the vendor's open-source packaging tool, generates an API server and deploys it to a cloud cluster.

Billing is usage-based, with no subscription tiers listed. Most public models are charged for the time they take to run, at a per-second rate set by the hardware. Some are charged per output instead. The pricing page lists flux-1.1-pro at $0.04 per output image, flux-dev at $0.025 per output image, flux-schnell at $3.00 per thousand output images and wan-2.1-i2v-480p at $0.09 per second of output video. Hardware rates include $0.000225 per second for an Nvidia T4 GPU. Larger multi-GPU A100 and H200 capacity depends on committed spend contracts.

The service scales with traffic, and it offers metrics and logs. An enterprise route adds an account manager, priority support, higher GPU limits, performance SLAs and volume discounts.

• Cloud API for running models
• Community model library
• Fine-tuning with your own data
• Custom model deployment with Cog
• Automatic scaling
• Metrics and logs
• Node, Python and HTTP examples
• Enterprise support and volume discounts

Our take

Replicate fits developers and product teams who want model-powered features behind an API, or who want to host a custom model without managing GPU infrastructure. Usage-based billing suits workloads that rise and fall, but the cost depends on the model, the hardware and the run time, so each model page needs checking for estimates. The pages do not describe a free plan or trial length, and no fixed subscription is listed. Private models on dedicated hardware accrue charges while idle, which matters for low-traffic projects. People who want a point-and-click app rather than code will find the entry point technical, since the pages center on API calls and code samples.