FLUX 3 pricing per second vs how Sume prices video

BFL's FLUX 3 Video is priced per second of output: QHD (2560x1440) at $0.40 per second and UHD (3840x2176) at $0.80, as of BFL's 2026-09-10 release note. Sume does not sell FLUX 3 Video, and its video models are priced by the SKUs GET /v1/videos/models returns, which are not always per-second, so a comparison starts with converting units.

BFL's numbers are from its release notes and Sume's from the Video generation docs, read 2026-09-29.

What does FLUX 3 Video cost on BFL's API?

Multiply the rate by clip seconds. A 20-second QHD clip at the listed rate is 20 x $0.40 = $8.00, and 20 seconds of UHD is 20 x $0.80 = $16.00. Continuation clips have their own rates.

How does Sume price a video model?

Each video model reports pricing_skus in the catalog. The docs' example response for seedance-2 shows a SKU named per-1000-video-tokens at 0.0154, so that model is not billed as a flat per-second rate. Sume's rate card describes Video Router prices as provider list rates times Sume margin, with per-model rates in the catalog.

Read the SKU for the model you use, then price a real request with a dry run rather than guessing: on hosted MCP, dry_run=true gives an admission and cost preview without submitting.

Can I compare the two directly?

Only for a specific request. Pick a model, resolution and length, get Sume's preview for it, and set that beside BFL's per-second figure for a similar clip. Different models produce different results, so a lower or higher number is not a quality claim either way.

What is the quickest way to get a Sume number?

Pick the model from GET /v1/videos/models, note its supported_durations and supported_resolutions, and preview the request. Over hosted MCP that is dry_run=true on generate_video. The preview is an admission and cost preview, so you learn the reservation before anything is submitted.

Then run a single short clip and read the cost recorded for it, which is a better guide than extrapolating from a rate.

Why does the unit matter?

A per-second rate scales linearly with clip length, so doubling the clip doubles the cost. A token-based SKU depends on how the model counts a request, which can depend on resolution as well as length. Reading BFL's per-second numbers next to a token-based SKU without converting gives a wrong ratio in either direction.

This is also why the tables in this post stop at listing figures and do not rank them.