fal launches FLUX 3 Max, claiming a 22-fold speed edge in AI video

fal released FLUX 3 Max on October 19, 2026, offering developers a video model designed to shorten the wait between creative revisions. Built with Black Forest Labs, it supports generation from text, images and keyframes. fal claims a 22-fold speed advantage over Seedance 2.5 at the same reported image-to-video quality rating.

Further training, not just faster hosting

FLUX 3 Max builds on Black Forest Labs’ FLUX 3 foundation model. fal Research handled specialized post-training—additional training after the main model was built—alongside reinforcement learning and changes to the systems that run it. The work targets visual fidelity, coherent motion and closer adherence to prompts, as well as shorter generation times.

That makes this more than an existing model offered through another service. In fal’s launch announcement, Black Forest Labs CEO Robin Rombach describes Max as a version fal customized for developers and enterprises on its platform. fal says the changes improve both output quality and speed, rather than trading one for the other.

This is the first iteration of Max. fal says further post-training is underway to improve visual quality, generation speed and overall performance. Those are development plans, not additional improvements available at launch.

Quality evidence and speed claims are different

The speed figures need a careful reading. fal’s release gives 10.95 seconds to generate a 10-second video, versus 277.80 seconds for Seedance 2.5. VentureBeat reports a different fal comparison figure: 12.45 seconds for Max, against the same Seedance time. Neither is a guarantee of customer completion times.

Both timings come from fal, not independent speed testing. The comparison described by VentureBeat uses 10-second, widescreen clips at 720p or 768p, depending on the model. fal’s documentation cautions that settings and operating conditions affect completion time. A higher reasoning setting, which changes prompt processing, also increases generation time.

The quality evidence is stronger, but narrower than an outright victory. fal cites a preliminary score of 1,139 Elo across 4,868 votes in Artificial Analysis Video Arena’s independent human-preference evaluation. Elo summarizes which model’s outputs people prefer in comparisons. The release acknowledges overlapping confidence intervals among the leading models.

VentureBeat’s October 19 leaderboard check puts Max at 1,138, tied with Dreamina Seedance 2.5 and two points behind MiniMax H3. Their confidence intervals overlap, so the scores do not establish a clear quality winner. That evaluation covers image-to-video generation with audio; it does not establish equivalent leadership for footage generated from text alone.

Three ways in, temporary launch prices

Developers can access Max through fal’s API, the interface software uses to request generations. The three routes provide different starting points:

• Text-to-video starts with a written description.
• Image-to-video starts with a source image and can accept an ending image.
• Keyframes-to-video accepts up to 10 images placed along the sequence.

Documented clips run from five to 20 seconds. Output is generated at 720p, with optional upscaling to 1080p—not native 1080p generation. Audio can be switched on or off, and formats include landscape, portrait and square compositions.

VentureBeat lists promotional pricing through November 7 at $0.0425 per output second for 720p and $0.0725 for upscaled 1080p. A 10-second clip therefore costs $0.425 or $0.725. Regular rates are $0.17 and $0.29 per second, bringing those same clips to $1.70 and $2.90 after the discount.

Faster generation could reduce waiting, but it could also encourage more attempts. Ten versions of a 10-second 720p shot cost $4.25 during the promotion and $17 at regular rates, before staff time or editing. The useful business test is whether teams reach an acceptable clip sooner, rather than simply producing more variations.