See what Replicate costs in 2026, how its per-second GPU pricing works, and whether this open-source AI model hosting platform fits your project.
Replicate is an AI infrastructure company that hosts thousands of open-source machine learning models and exposes them through a unified API, so developers can generate images, video, speech, and text without setting up their own GPU servers. Founded in 2019 by Ben Firshman and Andreas Jansson and based in San Francisco, the company built its reputation on Cog, an open-source packaging tool that lets any model be containerized and deployed to Replicate's infrastructure with predictable versioning.
In November 2025, Cloudflare announced an agreement to acquire Replicate in order to integrate its catalog of 50,000+ AI models into the Cloudflare Workers AI platform; the acquisition closed in December 2025. Replicate continues to run as its own product and brand within Cloudflare as of mid-2026.
Replicate's core feature is one-line API access to a large catalog of community and proprietary open-source models spanning image generation, video generation, audio/speech, upscaling, and large language models, with each model versioned so outputs stay reproducible over time. Developers who want to deploy their own custom model can use Cog to package code and dependencies into a standard container that Replicate can serve automatically on the appropriate GPU hardware.
Additional platform capabilities include webhooks for asynchronous predictions, fine-tuning support for select model families, private deployments with dedicated hardware, and an official Python and Node.js SDK for integrating predictions directly into application code.
Replicate has no monthly subscription; every account pays strictly for the compute it consumes, billed per second of hardware time. CPU instances start at roughly $0.000025/sec, while GPU rates range from about $0.000225/sec for an Nvidia T4 up to $0.001525/sec for a single Nvidia H100, with multi-GPU configurations priced proportionally higher.
Public community models bill only for active prediction time, while private models and standing deployments also bill for idle and setup time since the hardware stays reserved. Certain high-traffic model categories, especially large language models and generative media models, instead bill per token or per second/frame of output. Organizations with large or predictable workloads can negotiate custom enterprise pricing with volume discounts and committed-spend contracts.
Replicate lets developers run and deploy open-source machine learning models, such as image generators, video generators, speech models, and large language models, through a simple API without managing their own GPU servers.
Replicate has no monthly subscription. Customers pay per second of compute used, with rates starting around $0.000025/sec for CPU and going up to about $0.001525/sec for a single Nvidia H100 GPU; some models bill per token or per output unit instead.
Replicate was founded in 2019 by Ben Firshman and Andreas Jansson, and is headquartered in San Francisco, California.
No. Cloudflare announced its acquisition of Replicate in November 2025, and the deal closed in December 2025. Replicate continues to operate under its own brand within Cloudflare.
Cog is Replicate's open-source tool for packaging machine learning models and their dependencies into standard containers that can be deployed and versioned on Replicate's infrastructure.
Replicate does not advertise a dedicated free subscription tier; new accounts typically pay as they go for compute used, and enterprise customers can negotiate committed-spend contracts.
Yes. Developers can package a custom or fine-tuned model with Cog and push it to Replicate, where it can be served through the same API used for community models.
Yes, Replicate supports fine-tuning for select model families, allowing users to train models on their own data before deploying them.