Replicate

Replicate runs machine learning models behind a hosted API, so developers can call image, video, audio and language models without provisioning GPUs. The platform covers three jobs: running existing models, fine tuning them on custom data, and deploying private models of your own. Calls are made from Node, Python or plain HTTP, and the vendor's framing is that importing an image generator should feel like importing a package. The catalog mixes open source community models with official proprietary ones, spanning image generation and editing, vector output, speech, music, image restoration, image to video, captioning and large language models. Private models are packaged with Cog, the company's open source tool for containerizing machine learning models, and most of them run on dedicated hardware rather than a shared queue. There is also an enterprise track for teams that need committed capacity. Pricing is pay as you go with no subscription. Public models are billed either by run time at a per second hardware rate or by input and output, and each model page carries an estimate. Published examples include $0.025 and $0.04 per output image for image models, $3.00 per thousand output images for a fast image model, $3.75 per million input tokens with $0.01 per thousand output tokens for a reasoning model, and $0.09 to $0.25 per second of output video. Hardware rates run from $0.000025 per second, or $0.09 per hour, for a small CPU instance up to $5.04 per hour for an 80GB A100 and $5.49 per hour for an H100, with multiple GPU configurations available under committed spend contracts. Private models accrue charges for setup and idle time as well as active processing. One notable detail: Replicate's own about page now directs job seekers to the Cloudflare careers page.