fal solved a real problem. Running generative media models yourself means GPUs, cold starts, queueing and a great deal of infrastructure work that has nothing to do with your product, and fal removes all of it behind a consistent API across a catalogue of more than a thousand models. The documentation is organised around that proposition in three clear tracks, model APIs for calling something hosted, serverless for running your own code, and compute for dedicated GPU capacity, each with its own quickstart. That structure is genuinely helpful because those are three different jobs and most platforms blur them.
The sandbox deserves specific credit. Generative media output is not something you can evaluate from a description, and being able to try a model in the browser before writing a line of integration code saves an enormous amount of wasted work. The code samples across Python, JavaScript and cURL are consistent, the deployment and observability guidance is more thorough than most inference vendors bother with, and the blog carries reasonable tutorial material alongside model releases. The company has raised aggressively through 2025 and into 2026 at rapidly climbing valuations, and the platform has the polish that buys.
Two real gaps. The first is cost, and it is the one that will bite people. Billing spans three different models depending on which product you use, per request for model APIs, per second of execution for serverless, and per hour for dedicated compute. Each is reasonable on its own and the combination is genuinely hard to reason about, particularly for generative media where a single video request can cost orders of magnitude more than an image.
There is no cost estimation guidance, no worked example of what a realistic workload actually bills, and no advice on which billing model suits which pattern. That is the single most useful thing this documentation could add and it is missing. The second gap is choice. A catalogue of a thousand models is a wonderful asset and a genuine burden, because a newcomer has no basis for picking between them.
There is no comparison material, no guidance on tradeoffs between quality, speed and price for a given task, and no opinionated starting recommendation. The catalogue is a list, and lists do not help people who do not already know what they are looking for. I would also want something on licensing. These models carry a wide range of terms, some restrict commercial use, and a platform that makes a thousand of them equally easy to call has some responsibility to flag which ones you can actually ship with.
It does not. Three point four. Clean, well structured documentation for genuinely excellent infrastructure, held back by no on ramp, no help choosing, and complete silence on what any of it costs.