fal

What is fal?

fal is a generative media cloud that gives developers hosted APIs for image, video, audio and 3D models, plus serverless GPUs for running their own models. It is made by fal and is used to add model-generated media to products without building a GPU cluster or managing inference infrastructure. The problem it solves is the operational gap between a model demo and a production feature: fal handles the endpoint, scaling and billing so a team can call a model like any other API.

Key features

• Model APIs. A catalogue of image, video, audio and 3D models is available through hosted endpoints, including FLUX, Kling, Seedance, Nano Banana and MiniMax speech models.
• Per-output pricing. Most model APIs are billed per image, per second of video, per character of speech or per unit of compute, so cost tracks usage rather than a fixed seat count.
• Serverless GPUs. Teams can deploy their own models and pipelines on GPUs such as the H100, H200, B200 and B300, billed per second while the runner is alive.
• Queue and scale. The platform queues requests during traffic spikes and scales endpoints without the team provisioning servers by hand.
• Custom deployments. Dedicated deployments and enterprise support are available for teams that need reserved capacity or custom models.
• Simple client libraries. Official Python and JavaScript clients call an endpoint by its model ID and return a typed result.

Who is fal for?

Product and engineering teams adding generative media to an app, such as a design tool, marketing platform, video editor or game studio. It suits AI engineers who want to compare and swap models without managing infrastructure, startups that need usage-based pricing before committing to GPUs, and media companies running batch image or video jobs. Teams that need dedicated capacity or custom model hosting can move to serverless GPUs or a custom deployment.

Use cases

• Generating product images or marketing visuals from a text prompt.
• Creating short video clips from a still image for social campaigns.
• Adding text to speech or music generation to a creative app.
• Removing backgrounds or upscaling images in a batch pipeline.

Pricing

fal uses prepaid credits for model APIs and bills serverless GPUs per second. Model prices vary by endpoint: FLUX.1 schnell costs $0.003 per megapixel, Nano Banana costs $0.0398 per image, and Kling Video v3 Pro costs $0.14 per second. Serverless GPU list prices include $4.50 per hour for an H100, $6.00 per hour for an H200, $7.99 per hour for a B200 and $12.99 per hour for a B300, with lower rates for longer commitments. See the official fal pricing page.

Getting started

1. Create a fal account and generate an API key from the dashboard.
2. Set the key as FAL_KEY in your environment.

pip install fal

import fal_client
 
result = fal_client.subscribe(
    "fal-ai/flux/schnell",
    arguments={"prompt": "A developer workspace at night, soft light"},
)
print(result)

Alternatives

• Replicate runs open source models through a hosted API.
• Runpod provides GPU cloud instances and serverless endpoints.
• Baseten hosts and scales machine learning models in production.

FAQ

Is fal free?

fal does not advertise a free tier on its pricing page. Model APIs are paid through prepaid credits, and serverless GPUs are billed per second. New users can check the dashboard for any current credit offer.

What is fal used for?

fal is used to run generative media models in production. Teams call hosted APIs for image, video, audio and 3D generation, or deploy their own models on serverless GPUs.

How much does fal cost?

Model APIs are priced per output, such as $0.003 per megapixel for FLUX.1 schnell or $0.14 per second for Kling Video v3 Pro. Serverless GPUs start at $4.50 per hour for an H100.

fal vs Replicate?

Both host generative models behind an API. fal emphasises media models and serverless GPU deployments with per-second billing, while Replicate offers a broad model catalogue and a community of published models. The better choice depends on which models you need and how much control you want over the serving layer.

Yes. fal Serverless lets you deploy custom models and pipelines on GPUs such as the H100, H200, B200 and B300, billed per second while the runner is active.