fal Launches FLUX 3 Max, a Post-Trained Video Model Benchmarked at About 12 Seconds per 10-Second Clip

fal launched FLUX 3 Max, a post-trained version of Black Forest Labs' FLUX 3 that generates a 10-second video in about 12 seconds, according to fal's FLUX 3 Max page. fal says that speed lets you try twenty takes of a shot in the time Seedance 2.5 returns one.

• A preliminary tie at the top. fal reports a 1,139 Elo score in Artificial Analysis's preliminary image-to-video arena, which it describes as tied for #1.
• One model for text and image input. fal says text-to-video and image-to-video share the same post-trained model and the same 720p price.
• Start frame, end frame, up to 20 seconds. The image-to-video endpoint takes a starting image and prompt, an optional ending frame, and durations from 5 to 20 seconds.
• 75% off until November 7. 720p output costs $0.0425 per second during the promotion, against a list price of $0.17 per second.
• One API with H3 Max. fal says a shot can move between FLUX 3 Max and its post-trained H3 Max by changing a single model ID.

fal's numbers put FLUX 3 Max level with Seedance 2.5 on quality at a fraction of the wait

fal lists FLUX 3 Max and Seedance 2.5 at the same 1,139 Elo in Artificial Analysis's AA-Video-I2V v1.1 arena, with FLUX 3 Max finishing a 10-second clip about 22 times faster. The company labels the result preliminary and notes that MiniMax H3 (768p) sits first on that board at 1,141, with FLUX 3 Max's confidence interval overlapping the leaders.

Here is fal's own comparison, with times averaged for a 10-second, 16:9 video:

Model	Image-to-video Elo	Time for a 10s video
FLUX 3 Max	1,139	12.45s
Gemini Omni Flash 1.1	1,134	47.37s
MiniMax H3 (768p)	1,141	83.26s
FLUX 3	1,089	126.44s
Wan 3.0	1,127	180.16s
Seedance 2.5	1,139	277.80s

The public Artificial Analysis image-to-video leaderboard ranks models by Elo from pairwise human preference votes. We could not confirm FLUX 3 Max's placement from its public listing, and its figures for some models differ slightly from fal's table, such as 1,140 for MiniMax H3 (768p) and 1,132 for Gemini Omni Flash 1.1.

fal's own documentation cautions that rankings can change as evaluations receive more votes. Its image-to-video page also cites a benchmark figure of approximately 11 seconds and says actual generation time varies with request settings and operating conditions.

The image-to-video endpoint directs motion from a starting frame

On the image-to-video endpoint, the starting image sets the opening composition while the text prompt directs subject movement, camera motion and changes within the scene.

• A start frame and prompt are required. An optional ending frame guides where the sequence finishes, and images can be PNG, JPEG or WebP.
• Clips run 5 to 20 seconds. Duration moves in one-second increments, with 5 seconds as the default.
• 720p is the default. Optional 1080p upscaling is available for higher-resolution output.
• Audio is generated by default. It can be switched off.

The endpoint also offers aspect ratios from 21:9 to 9:16, plus a reasoning level and an optional seed.

A matching text-to-video endpoint handles scenes without a starting image. fal says both modes share the same post-trained model.

Pricing drops to $0.0425 per second at 720p until November 7

fal bills FLUX 3 Max per second of output video, with no minimums and no subscription. The 75% discount is temporary, and the main product page lists 720p at $0.0425 per second during the promotion and $0.17 per second after November 7, for both text-to-video and image-to-video.

At those rates, a 10-second 720p clip costs about $0.43 during the discount and $1.70 at list price.

The image-to-video endpoint page also lists optional 1080p upscaling at $0.0725 per second during the promotion and $0.29 per second after.

fal says content generated through its API can be used in commercial projects, subject to its terms of service.

FLUX 3 Max sits alongside H3 Max, and fal is still training it

fal Research worked on three layers at once on top of Black Forest Labs' base model. Post-training with Black Forest Labs targeted visual fidelity, motion coherence, prompt adherence and image-to-video quality, while inference optimization cut a 10-second clip from 126.44 seconds on FLUX 3 to 12.45 seconds.

Systems engineering tuned the serving path on fal's own inference engine. fal says it is continuing to post-train the model.

The base FLUX 3 keeps its own draft, keyframe and extend endpoints. fal's FAQ lists text-to-video and image-to-video for FLUX 3 Max at launch, though its image-to-video page also links a keyframes-to-video endpoint.

The same team built H3 Max, fal's post-trained version of MiniMax's H3 model. In fal's comparison, H3 Max is the only faster model at 7.55 seconds for a 10-second clip, and both run on the same fal API.