FLUX 3 Max makes a 10-second AI video in about 12 seconds, and it is 75% off for three weeks

fal Research and Black Forest Labs have launched FLUX 3 Max, and the headline number is the one every AI filmmaker will notice first. A 10-second clip at 720/768p in about 12 seconds. That is not a preview render. @fal says it is tied for number one in image-to-video quality on independent third-party benchmarks.

The second number is the price. For the next three weeks FLUX 3 Max is 75% off on fal, which the company calls the most affordable frontier video model at its quality level. If you have ever burned an afternoon waiting on renders to test one camera move, this is aimed squarely at you.

What the companies actually claimed

Per @bfl_ai, FLUX 3 is a model designed for customisation, and fal adapted and optimised it to create FLUX 3 Max. fal describes it as post-trained by fal Research in partnership with Black Forest Labs. BFL's own launch post says it is 22 times faster than Seedance 2.5 at the same quality.

The benchmark detail comes from a creator thread. @ProCreatorHarry cites a 1,139 Elo score on Artificial Analysis and a 50-point Elo gain over the original FLUX 3, attributed to fal's benchmarks. fal also says FLUX 3 Max joins its H3 Max at the frontier of video generation, and that it excels at photorealism, with lifelike textures, natural motion and cinematic lighting.

What was not said

There is no stated price per second or per clip, only the 75% discount. We do not know what the full price will be after three weeks, so any budget you plan around the promotional rate is a budget for three weeks only.

The quality claim is specific to image-to-video. The 22 times figure comes from the companies themselves, and the same-quality comparison with Seedance 2.5 is a ranking on one benchmark, not a verdict on every kind of shot. Elo ties also hide a lot. Two models can sit level overall and still differ sharply on hands, crowds, fast camera moves or long takes.

Output is 720/768p. That is a draft and social resolution, not a finishing one. Nothing in the announcements mentions audio, clip length beyond 10 seconds, or reference controls. Treat those as unknown until you have tried it.

Where it fits in a real workflow

Speed changes how you work before it changes how good the work is. At 12 seconds a clip, you can run a dozen takes of a shot in the time a slower model returns one. That suits blocking, previs, animatic passes and look development, where you want volume and judgement, not a single polished render.

It also changes the economics of failure. Cheap, fast generations make it reasonable to throw away nine out of ten, which is how most of us end up finding the good one anyway.

There is early evidence of what quick generation enables. @BlendiByl shared an open-source flow in which FLUX 3 Max generates the video, GVHMR extracts 3D motion from it on a single GPU, and the result is retargeted live onto a Mixamo character. That is one creator's pipeline, not a benchmark, but it shows the point. When generation is nearly real time, video becomes an input to other tools, not a final deliverable you wait for.

What to test first

Start with your hardest image-to-video shots, not the easy ones. Take three stills you have already struggled with, ideally a face turning, a hand interacting with an object and a moving vehicle, and run them through FLUX 3 Max and your current model side by side.

Then check consistency across a sequence, not a single clip. Generate five shots of the same character from the same reference and see whether they cut together. Speed is only useful if the keepers survive contact with an edit.

Finally, use the discount window deliberately. Run your tests inside the three weeks, log what you spent, and decide on the full price once it is announced. If the quality holds at 720p, the sensible use is fast drafts on FLUX 3 Max, with a slower model reserved for the hero shots.

Sources

- fal's launch announcement
- fal's introduction of FLUX 3 Max
- Black Forest Labs' launch post
- fal's post on speed and latency
- Benchmark and Elo summary from @ProCreatorHarry
- Real-time animation pipeline demo from @BlendiByl