One API in front of 1,500+ generative media models, plus GPUs for inference and training.
What it does
On this page
fal is infrastructure for generative media. A single API fronts a catalogue of more than 1,500 image, video, audio and 3D models, and the same account can rent GPUs for inference, fine-tuning and training.
Who it is for
Developers and product teams shipping media generation inside an application who would rather not sign a contract with every model vendor.
How it works
Calls go to named model endpoints, so swapping a model becomes a change of path rather than a new integration. The company runs its own inference engine for diffusion workloads and autoscales serverless deployments from idle to large fleets; custom weights can be deployed privately. LoRA training is offered as an endpoint, and an MCP server exposes the models to compatible assistants. Billing follows output, so a text-to-video model is priced per second of generated footage, while GPU rental is hourly.
Pricing
Model APIs bill per output unit. GPU list prices start near $4.50 an hour for an H100, advertised as low as $2.49 with commitments, and rise for H200, B200 and B300 cards. Enterprise arrangements go through sales.
Notable limits
There is no subscription tier, so cost scales directly with generation volume and GPU time: cheap to test, expensive to leave unmonitored.
Source: https://fal.ai/