fal is a generative media platform for developers. It serves over 1,000 image, video, audio and 3D models through one API, and handles the GPU provisioning, autoscaling, model loading and performance tuning underneath. Switching from one model to another means changing a single identifier rather than rebuilding an integration.
Burkay Gur and Gorkem Yurtseven founded the company in 2021, originally as Features and Labels, and it is headquartered in San Francisco. Gur left Coinbase at the end of 2020 to start it. New models are typically available on the platform the day they are released.
Three calling methods are supported: a queue for long jobs, which is the default, synchronous calls for lowest latency, and WebSockets for interactive use. Workflows chain several models behind a single endpoint. The fal Inference Engine uses templated CUDA kernels, and FlashPack, a PyTorch model-loading format released in October 2025, is claimed to load models three to six times faster than existing methods.
Pricing works three ways. Serverless and dedicated compute are billed by GPU time, with H100 reported from 1.90 US dollars an hour against a 4.50 list rate and B300 from 4.50 against 8.50. Model APIs bill per output unit, so video is priced per second or per video and images per image or per megapixel. fal Agent, a conversational interface that picks and chains models, runs as a subscription at 50, 200 and 1,000 US dollars a month across its Starter, Pro and Max tiers, with negotiated enterprise contracts above that.