Replicate

Replicate is a developer platform for running AI models through an API. It is best known for image and video generation models, but the catalog spans text, audio, vision, upscaling, segmentation, 3D, and custom deployments.

The product sits between a playground and a full cloud stack. Developers can run official models with stable APIs, call community models, or publish their own model containers without operating GPU infrastructure directly.

When to pick Replicate

• You want to test a model quickly. Browser playground plus API examples make evaluation fast.
• Your app needs image or video generation. The catalog is deep and changes quickly.
• You prefer official model stability. Official models avoid version surprises and cold-boot pain.
• You need custom model deployment without DevOps. Package the model and let Replicate handle serving.
• You are comparing alternatives. Replicate is useful as a neutral test bench before committing to self-hosting.

When to pick something else

• Fastest media inference: Fal.ai is usually the speed-first pick for production image/video APIs.
• Open LLM infrastructure: Together AI or Fireworks AI.
• Serverless GPU apps beyond model calls: Modal gives more control over Python apps, jobs, and web endpoints.
• Creator workflows: Midjourney, Runway, Krea, or Leonardo.

Pricing

Replicate uses two main pricing patterns. Some public models are billed by input and output, such as images, video seconds, or tokens. Many public and community models are billed by the hardware used and the time they take to run.

Private models are different. Most private models run on dedicated hardware, so teams can pay for setup time, idle time, and active processing time while the deployment is online. Fast-booting fine-tunes are an exception when labeled that way. This is fair for experimentation but needs monitoring in production. Slow models, high-resolution outputs, idle deployments, and retries can move the bill faster than a flat SaaS plan.

As verified on 2026-06-25, Replicate’s hardware rates run from CPU at $0.09/hr (small) or $0.36/hr (standard), through T4 at $0.81/hr, L40S at $3.51/hr (2x at $7.02), A100 80GB at $5.04/hr (2x at $10.08), and H100 at $5.49/hr (2x at $10.98). 4x and 8x GPU configurations require committed-spend contracts. Output-priced examples include FLUX 1.1 Pro at $0.04 per image and Wan 2.1 i2v at $0.09 per second of 480p output video or $0.25 per second of 720p output video. Text-model examples include Claude 3.7 Sonnet at $3 per million input tokens plus $0.015 per thousand output tokens, and DeepSeek R1 at $3.75 per million input tokens plus $0.01 per thousand output tokens. Enterprise and volume-discount conversations can add higher GPU limits, performance SLAs, priority support, onboarding help, and custom-model optimization.

Evaluation checklist

Before using Replicate in production:

• Prefer official models when API stability and predictable pricing matter.
• Check the specific model page for cost estimates before building a feature around it.
• Measure cold starts, runtime, queueing, and retries on realistic prompts.
• Decide whether the model needs private deployment or can run as a public model call.
• Track high-resolution media, long video outputs, and failed runs separately.
• Compare custom deployments against Modal, Fal.ai, Together AI, Fireworks AI, or direct cloud GPUs once volume is predictable.

Buyer fit

Replicate is strongest for teams that are still exploring model choice. It lets developers compare image, video, audio, vision, and utility models quickly, then turn the winning model into an API call without building infrastructure first.

providers or self-managed GPU infrastructure may offer better latency, controls, or unit economics. Replicate is often the right first production path, but not always the cheapest final path.