Know the unit cost before you build.

Fetch GET /dev/pricing; it is public and returns data.version plus every data.prices[feature] entry. Use each entry's preferred usd value for normal charges. Keep multiplier units as multipliers, match the exact feature key and billing unit before quoting a cost, and respect the response cache headers.

The Unit column is supplied by the pricing catalog. Depending on the feature, a unit may be a use, output, image, reference, or second. Reference-to-video rows say per input + output second, because there a second of the reference video is billed just like a second of the result; GET /dev/pricing reports those keys as per_second.

Reference-to-video tiers charge for the reference video as well as the clip they produce, so a 5-second clip built from a 5-second reference bills 10 seconds, not 5. Video references make a generation more expensive than the image-to-video row above, not cheaper, even though the per-second number is lower. It applies to Seedance 2.0 and 2.5 with video references and to Wan 2.6 R2V. fal.ai advertises the Seedance tier as a 0.6× multiplier, which reads like a discount and is not one: over a duration that now includes the input, 0.6 × (input + output) works out to about 1.2× what the same clip costs from a still image. Those Seedance rows compare like for like, because fal.ai bills them on the same input + output basis; on Wan 2.6 R2V, fal.ai states only a flat per-second rate, and the tooltip on its price says so.

Your balance and feature prices use one fixed dollar rate, so the displayed unit price does not change with the amount you add.

Our billing records distinguish successful consumption from requests that fail before usable processing occurs. See the Refund Policy for the applicable rules.