FLUX 3 Max: The AI Video Model Fast Enough to Direct by Revision, at a Price That Quadruples on November 7
It ties the best video models on quality and beats them badly on waiting time. That changes how you work more than any quality bump would.
Most AI video tools make you wait about as long as a coffee break for ten seconds of footage. Four or five minutes per take is normal, and it shapes everything about how people use these tools. You write a long, careful prompt, you hit generate, you go do something else, and you come back hoping. It is closer to placing a bet than directing a shot.
FLUX 3 Max, the video model fal launched in October with Black Forest Labs, comes back with a ten-second clip in about twelve seconds, according to fal's own timing. That is roughly the time it takes to reread your prompt.
The quality is not the story. On the leaderboard everyone cites, it ties. The story is what happens to your process when a take costs less time than thinking about the take. And there is a date attached to all of this that anyone planning real work should circle.
What FLUX 3 Max is, in plain terms
Black Forest Labs makes the FLUX image and video models. fal is a company that hosts AI models and sells access to them by the second. For FLUX 3 Max, the two split the work. fal's launch page puts it this way: "Black Forest Labs brought the foundation model; fal Research brought post-training." In other words, BFL built the base video model, and fal trained it further to follow instructions more closely and run faster on its own servers.
What you can make with it:
- Clips from 5 to 20 seconds long, from a text prompt.
- Clips that start from your own still image, with an optional final frame so the shot lands where you want it.
- Clips guided by several keyframes, still images the clip must pass through along the way (VentureBeat reports up to 10 images in that mode).
- Sound generated with the picture, which you can switch off.
- Shapes from ultra-wide 21:9 down to vertical 9:16, at 720p, or 1080p for a higher price.
You cannot download it. Nothing in fal's materials mentions released model files, so this is a hosted service: you send a request, you pay per second of finished video, you get a file back. fal's page says content made through its API "can be used in commercial projects," with the details in fal's terms of service.
It is also showing up inside other tools. Krea added it on October 20, and ComfyUI lists it among its partner nodes, so you may not need a fal account to try it.
The speed claim, and how much to trust it
fal's numbers: a ten-second clip in 12.45 seconds, against 277.80 seconds for ByteDance's Seedance 2.5 and 126.44 for the original FLUX 3. That is about 22 times faster than Seedance and 10 times faster than its own parent model.
Read those as fal's averages, not a promise. VentureBeat, which covered the launch on October 19, notes they come from ten-second, widescreen clips at 720p or 768p, and that settings and conditions change the result. Your clip at 1080p, or with the slower "high" reasoning setting (where the model rewrites your prompt before it starts), may take longer. And fal's own FAQ concedes that one other model, H3 Max, is faster still at 7.55 seconds, though it is left out of fal's comparison table. Fast is a category here, not a crown.
Even with those caveats, the gap is large enough to matter. Twelve seconds versus four and a half minutes is the difference between iterating and waiting.
The quality claim, and why "tied" is the honest word
On Artificial Analysis's image-to-video leaderboard, the public ranking built from people voting between pairs of clips, VentureBeat found FLUX 3 Max at 1,138 points on October 19. Seedance 2.5 sat at the same score. MiniMax H3 was two points ahead.
Two points is nothing. The leaderboard's own uncertainty ranges overlap, and VentureBeat reports the model's possible rank as anywhere from first to fifth. That ranking also covers starting from an image with sound on. It says nothing about pure text-to-video.
So the fair summary is this: FLUX 3 Max looks as good as the best hosted video models right now, by public vote, and nobody can say it looks better. If someone sells it to you as "the new number one," they are rounding up.
Why the speed matters more than the score
Here is my position. A tie on quality plus a twenty-fold cut in waiting time is a bigger deal for working creators than a small quality win would have been.
When each take costs four minutes, you protect yourself by front-loading the thinking. Long prompts, every detail specified, minimal attempts. You get good at writing, not at directing.
When each take costs twelve seconds, that logic flips. The cheapest way to find a good shot becomes trying a rough version, looking at it, and changing one thing. It is how you would work with a camera operator on set: "Same move, slower. Again, but start wider." The craft moves from prompt engineering to giving notes.
Two settings make this even more practical:
Reasoning level. The default is "high," and fal says the model expands your prompt. That is useful when you write three words and want a scene. It is a problem when you are giving precise notes, because the model is rewriting them before it starts. Try "low" when you want it to do exactly what you said.
Seed. A seed is the number that fixes the random starting point. Keep it the same between takes and you change one variable at a time, which is the whole point of giving notes.
The date to circle: November 7
Current pricing is a launch discount, 75 percent off, running until November 7.
| Now | After November 7 | |
|---|---|---|
| 720p, per second | $0.0425 | $0.17 |
| 1080p, per second | $0.0725 | $0.29 |
| One 10-second 720p clip | about $0.43 | $1.70 |
Forty-three cents a take makes revision feel free. At $1.70 it is still cheap next to a reshoot, but it adds up fast. Forty attempts to nail one ten-second shot is $17 today and $68 in December. VentureBeat makes the same point from the other side: faster generation can encourage more attempts, which can raise total spend unless more of those attempts are usable. Their cost comparison also leaves out your time and editing, which is usually the bigger bill.
So the real risk is building your habits and your client quotes on the launch price.
Put this into practice
Here is a routine for a short product or mood spot, built to take advantage of the speed while the discount lasts and to survive the price change after it.
1. Start from a frame you control. Use the image-to-video mode with a still you made or approved: a product shot, a location photo, a frame from your storyboard. Add an end frame if the shot needs to land on something specific, like a logo or a hero angle. You are giving the model less to invent.
2. Prompt the change, not the picture. fal's own tip for image-to-video is to describe what happens after the opening frame (the subject's action, the camera move) instead of re-describing everything already visible. fal's own example, "The camera slowly moves closer as the woman turns toward the window," beats a paragraph about her outfit.
3. Draft short, at 720p, with sound off. Five seconds, 720p, audio off, reasoning low, a fixed seed. You are checking blocking and camera movement, nothing else. Each of these drafts costs about 21 cents at the launch rate.
4. Give one note per take. Change one thing: speed of the camera, where the subject ends up, the light. Keep the seed. When a take works, write down the exact prompt and seed next to the output file.
5. Render the keeper once, properly. Full length, sound on, 1080p if the job needs it. This is the only expensive render.
6. Count your takes this week. Look at how many drafts it took to land each shot, then multiply by the post-November price. That number, not the launch price, belongs in your quotes.
Where it breaks
You do not own the tool. No downloadable model means no running it offline, no guarantee it stays available, and no control if fal changes the price or the model. Keep your stills, prompts and seeds in your own folders.
"Tied for first" is not "best." Two points on a voting leaderboard is noise. Test it against whatever you use now, on your own kind of shot, before you switch.
The default setting rewrites you. The "high" reasoning level expands your prompt. Great for exploring, bad for precise direction.
Resolution is modest. It generates at 720p. VentureBeat describes the pricier 1080p option as upscaling, so treat it as an enlargement, not extra detail from the model.
The speed numbers are the seller's. fal timed them, on its preferred settings. Expect your own runs to vary.
The price is temporary. Everything that feels effortless this month costs four times more after November 7.
What to do this week
The tools that change creative work are usually not the ones that score a point higher. They are the ones that change how often you can try something. FLUX 3 Max is the first hosted video model where trying again is nearly as fast as deciding to.
Use the discount window to learn the revision habit: short drafts, one note per take, a fixed seed, a single proper render. Then do the arithmetic at the real price before you quote anyone. If the habit cuts your takes per shot, the speed pays for itself. If it does not, you found out for the price of a few lunches.