Speed is the real story here, not the leaderboard. fal.ai launched FLUX 3 Max, a video model developed with Black Forest Labs, and says a 10-second clip takes 12.45 seconds to generate. Seedance 2.5 takes 277.80 seconds in fal’s own comparison. That is roughly 22 times faster at the same quality score.
If the numbers hold up in production, generated video starts to feel interactive. A designer can try a camera move, look at it, and revise while the idea is still fresh. Most video models still make you wait long enough to lose the thread.
A top score, with caveats
On the Artificial Analysis image-to-video leaderboard, checked October 19, FLUX 3 Max sits at 1,138 Elo. That matches Dreamina Seedance 2.5 and trails MiniMax H3 by two points (1,140). The confidence intervals overlap heavily, so nobody has a decisive lead. Artificial Analysis gives FLUX 3 Max a possible rank anywhere from first to fifth.
The test also covers image-to-video with audio only. It says nothing definitive about text-to-video. fal’s launch materials used a preliminary snapshot with slightly different scores and called it shared first place. The fair reading is simpler: FLUX 3 Max competes with the best models on this test.
What the speed claim does and doesn’t tell you
The timing is an average for 10-second, 16:9 clips at 720p or 768p, depending on the model. It’s vendor-reported, and the documentation admits that settings and operating conditions change completion time. The text-to-video endpoint has low and high reasoning modes, and the high setting is slower. That’s one more variable to test.
For developers building video into an app, a demo proves what’s possible. Load testing proves whether the speed survives real traffic.
What fal adds to FLUX 3 Video
FLUX 3 Max builds on Black Forest Labs’ FLUX 3 Video. fal says its extra training targets instruction following, visual quality and shorter generation time. That matters commercially. Inference platforms like fal are competing on more than hosting, and tuning a partner’s model is one way to stand apart from other API providers.
Black Forest Labs points to speech, physical realism, sound effects and odd scenes as strengths. Those are the developer’s claims, and the preference ranking doesn’t verify them individually.
- Public text-to-video and image-to-video endpoints, with optional end image on the latter
- Output from 5 to 20 seconds
- 720p generation, with optional upscaling to 1080p
- Audio on or off
- Landscape, portrait and square formats
- A keyframes endpoint that accepts up to 10 images along the sequence
The keyframes option is the most interesting for product demos and storyboards, where you want control over the stages a clip passes through.
Pricing
A 75% promotional discount runs through November 7. At 720p, that’s $0.0425 per second, or $0.425 for 10 seconds. Upscaled 1080p is $0.0725 per second, or $0.725. Regular prices are $0.17 and $0.29 per second, so $1.70 and $2.90 for 10 seconds.
fal says API output can be used commercially under its terms. That doesn’t grant rights to download or run the model yourself, and the launch materials don’t address that.
The number that matters is cost per accepted clip. Ten 10-second attempts at 720p cost $4.25 on promo and $17 at regular rates, before staff time and editing. Faster generation can also mean more attempts. If acceptance rates don’t improve, teams just spend more on each decision, only sooner.
Does it matter?
Yes, mostly because of speed. Quality parity with Seedance 2.5 and MiniMax H3 is table stakes now. A 22x time advantage, even if it shrinks in practice, changes how creative teams can work. But test it with your own prompts and settings before believing the headline. The real question is how often fast output turns into footage worth keeping.
