fal’s FLUX 3 Max ties Seedance 2.5 and promises faster video

A tie is not a clear lead

In Artificial Analysis’ image-to-video leaderboard, checked on October 19, FLUX 3 Max scored 1,138 Elo points. Dreamina Seedance 2.5 scored the same; MiniMax H3 was two points ahead.

Artificial Analysis places FLUX 3 Max second by point estimate, but its possible rank spans first to fifth. The test measures viewers’ preferences between videos made from source images, with audio. It does not establish how well the model handles text-only prompts or every production requirement.

fal’s launch materials cite an earlier snapshot with slightly different scores and call the result a joint first, based on overlapping confidence intervals. The safer reading is that FLUX 3 Max is competing with the leaders, not that it has demonstrated a distinct quality advantage.

Speed with conditions attached

fal says it measured generation on 10-second videos in a 16:9 aspect ratio, at 720p or 768p depending on the model. FLUX 3 Max took 12.45 seconds—close to the length of the video it generated, but still slower than real time.

Those are vendor-published measurements, not a guarantee for every customer request. fal’s documentation says generation time varies with settings and operating conditions. Its text-to-video endpoint also offers low and high reasoning levels; the high setting takes longer.

For a team adding video generation to an application, a fast demo is a starting point. Production testing needs to establish whether that speed holds under the expected workload, and whether the slower reasoning setting changes the trade-off.

What fal adds to the base model

FLUX 3 Max is based on Black Forest Labs’ FLUX 3 Video. fal says additional training is intended to improve instruction following and video quality, while its work on the model also targets shorter generation times.

That is part of the value an API provider can offer beyond access to a model: additional training and execution improvements. Black Forest Labs describes speech, convincing physical behavior, sound effects and unusual scenes as strengths. Those are the developer’s assessments; the preference leaderboard does not independently verify each capability.

The API includes open endpoints for both text-to-video and image-to-video. Users can provide a prompt or a source image, and image-to-video can also take an image intended as the clip’s ending frame. Videos can run from five to 20 seconds, are generated at 720p, and can be upscaled to 1080p. Audio can be switched on or off, and landscape, portrait and square orientations are supported.

A separate keyframe endpoint accepts up to 10 images arranged along a sequence—useful for teams that need more control over the visual stages of a clip.

The cost of trying more often

fal’s published pricing includes a 75% discount through November 7. At the stated per-second rates, a 10-second video costs $0.17 at 720p or $0.29 at 1080p at the regular price. fal says API-generated content can be used commercially subject to its terms; that does not establish a right to download, modify or run the model independently.

At the launch rate, generating 10 versions of a 10-second 720p video costs $4.25; at the regular rate, it costs $17. Neither figure includes staff time or later editing.

I think the more useful test is not whether FLUX 3 Max generates a clip quickly, but how often a team gets one it actually keeps. Faster generation can reduce waiting—and encourage more attempts. If the share of acceptable results stays the same, a faster workflow could still mean a higher bill.