Replicate

Replicate is a hosted inference platform for machine-learning models: it runs thousands of open-source and proprietary models behind a single API and bills by the second of compute, so developers can ship AI features without provisioning or managing GPUs. The problem it solves is that the gap between an open-source model existing and that model serving production traffic is mostly infrastructure work - GPU procurement, CUDA versions, container builds, cold starts, autoscaling, and paying for idle capacity between requests. Replicate collapses that into an HTTP call. You pick a model, send input, get output, and pay only for the seconds it ran. Billing is genuinely usage-based, with two shapes depending on the model. Most public models are billed by hardware and runtime, with the price per second set by the GPU class in use, and every model page carries a cost estimate. Others are billed by input and output volume. Published examples as of 2026 include 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, Recraft V3 at $0.04 per output image, Ideogram v3 Quality at $0.09 per output image, DeepSeek R1 at $3.75 per million input tokens and $0.01 per thousand output tokens, and Claude 3.7 Sonnet at $3.00 per million input tokens and $0.015 per thousand output tokens. There is no monthly platform fee and no minimum commitment. Beyond running public models, Replicate supports fine-tuning models on your own data and deploying private custom models using Cog, its open-source containerisation tool for ML - which is the piece that makes the platform usable for work that is not just calling someone else's checkpoint. The two practical caveats: cold starts on infrequently used models are real and can dominate latency for a low-traffic app, and per-second pricing that looks trivially cheap at prototype scale can become the wrong cost structure at sustained high volume, where a dedicated GPU is cheaper. Who it is for: product engineers adding image generation, transcription, video, or LLM features to an existing app, and teams prototyping with several models before committing to one. How it compares: Fal.ai is the closest competitor and is generally faster for image and video workloads. Together AI and Fireworks focus on hosted LLM inference at lower per-token cost. Hugging Face Inference Endpoints gives dedicated managed instances rather than shared per-second billing. Modal is more general-purpose serverless compute and requires more setup for more control. Going direct to OpenAI or Anthropic is cheaper for their own models but gives no access to the open-source ecosystem. Use Replicate when model variety and zero infrastructure matter more than the last cent per inference.