Replicate

by Cloudflare • • Large Language Models (LLMs)

No reviews yet

Replicate is a cloud API for running, fine-tuning and deploying open-source and custom AI models, built for developers, billed by usage and owned by Cloudflare.

Replicate puts AI models behind one API, so an application can call one without anyone setting up GPU servers. Cloudflare announced the purchase on 2025-11-17, and Replicate has been owned by Cloudflare since 2025-12-01. The service keeps its own name, site and billing. Cloudflare said existing APIs and workflows would keep working without interruption. Each model has a page with a Playground, where inputs are filled in as a form and the same run appears as JSON, Node.js, Python or HTTP code. Explore and the collections group models by task, from image and video generation to language and embedding models. Official models keep a stable API and bill per output, while other models bill for the hardware time they use. Teams package their own models with Cog, Replicate's open-source tool, and publish them as public or private models. Fine-tuning trains a model such as FLUX on a team's own images, and webhooks report when an asynchronous prediction finishes. Replicate is built for software developers who want image, video, audio or text generation in an application without running GPUs themselves. It suits startups shipping a first AI feature, product teams testing several open models side by side, and machine learning engineers serving a custom model. Usage is paid from credit bought in advance, by the output for official models and by the second of hardware for the rest. Larger teams sign enterprise contracts for volume discounts, reserved compute and multi-GPU capacity. A typical job is a web app that sends a user's prompt to an image model and stores the picture that comes back. Developers make that call from the Python or Node.js client, often inside a Next.js app, and let a webhook hand the finished prediction to the rest of their code. Teams on Cloudflare can make the same calls from a Worker, following Replicate's guide for Cloudflare Workers. Other common jobs are captioning images, transcribing speech, generating music and restoring old photos. Each of them starts as a run in the Playground and ends as one line of code in production.

Pricing

View plans

Pay as you go

Custom pricing
- From $0.000025 a second (cpu-small, $0.09 an hour), usage-based, no billing toggle and no flat fee
- Credit is bought in advance, valid 1 year from purchase and not refundable; new work stops when the balance reaches zero
- Official models bill per output: flux-dev $0.025 an image, flux-schnell $3.00 per thousand images, recraft-v3 $0.04 an image
- ideogram-v3-quality $0.09 an image; wan-2.1-i2v video $0.09 a second of output at 480p and $0.25 at 720p
- Other models bill per second of hardware: cpu $0.0001, Nvidia T4 $0.000225, L40S $0.000975, A100 80GB $0.0014, H100 $0.001525
- 2x GPUs per second: L40S $0.00195, A100 80GB $0.0028
- H200 at any size, 2x H100 and 4x and 8x A100, H100 and L40S need a committed spend contract: from $0.001525 a second (H200) to $0.0122 (8x H100 or H200)
- Run public models, deploy private custom models and fine-tune models
- Models in the Try for free collection run at no cost for a limited number of runs; the billing docs say credit is bought before use
- Rate limit 600 prediction requests a minute, 3,000 a minute on other endpoints
- API prediction inputs, outputs, files and logs are removed after an hour by default
Enterprise

Custom pricing
- Price on request, contract terms
- Volume discounts and reserved compute with predictable costs
- Guaranteed uptime and performance
- Additional multi-GPU capacity through committed spend contracts