Models#Video generation#Benchmark
fal launches FLUX 3 Max video model at $0.0425 per second for 720p on promotion
fal launched FLUX 3 Max on October 19: 720p at a promotional $0.0425 per second through November 7 (list $0.17), and an independent board ranks it second.
fal launched FLUX 3 Max on October 19, 2026. It is a video generation model that fal post-trained, together with Black Forest Labs, from the FLUX 3 Video model, and it is sold as an API on fal’s inference platform. fal’s model page lists a promotional price of $0.0425 per second at 720p, against a list price of $0.17, with the discount running until November 7.
Key points
- Price: fal’s model page gives $0.0425 per second at 720p and $0.0725 at 1080p as promotional rates. The list prices are $0.17 and $0.29. The page dates the promotion from October 15 to November 7, while the press release says “the first three weeks.”
- Speed: fal’s own benchmark for a 10-second clip is 10.95 seconds in the press release and 12.45 seconds (720p) on its landing page. Against the 277.80 seconds the release gives for Seedance 2.5, those figures work out to roughly 25× and 22× faster.
- Ranking: On Artificial Analysis’s image-to-video board, which includes audio, FLUX 3 Max sits second at 1138 Elo, two points behind MiniMax H3 at 768p (1140). The intervals overlap. fal’s landing page says “tied for #1.”
Background
Black Forest Labs made FLUX 3 Video generally available on August 4, 2026, through its own API and select partners, according to its blog. FLUX 3 Max builds on that model; fal’s page calls it a “post-trained variant of FLUX 3 Video.” Black Forest Labs released the image model FLUX 3 Image on October 2, from the same model line. Four days before this launch, Kuaishou moved Kling 4.0 to full release.
The facts
Most figures below come from fal and Black Forest Labs. Only the leaderboard comes from a third party.
| Item | Figure | Source |
|---|---|---|
| FLUX 3 Max | A post-trained variant of FLUX 3 Video, built with Black Forest Labs | fal model page |
| 720p price, promo / list | $0.0425 / $0.17 per second | fal model page |
| 1080p price, promo / list | $0.0725 / $0.29 per second | fal model page |
| Duration | 5 to 20 seconds, 5 by default | fal model page |
| Audio | On by default, can be turned off | fal model page |
| 10-second generation | 10.95 s; about 11 s; 12.45 s (720p); 12 s | Release; model page; landing page; BFL page |
| Against Seedance 2.5 | 277.80 s; “22×” or “more than 25×” | Release, landing and BFL pages (22×); model page (more than 25×) |
| Image-to-video board | Second, 1138 Elo (±15), 5,054 samples | Artificial Analysis, read October 21 |
| Figures in the release | 1,139 Elo across 4,868 votes | fal release, October 19 |
| Launch modes | Text-to-video, image-to-video, keyframes-to-video | Release and model page; landing FAQ says “Two at launch” |
Several figures do not line up. The model page starts the promotion on October 15, four days before the launch, and “the first three weeks” counted from October 19 would end around November 9, not November 7. The landing page’s FAQ names two endpoints at launch, while the release and the model page list three. The Elo numbers drift as well: the release says 1,139 across 4,868 votes, and the board now shows 1138 across 5,054 samples. The board has no snapshot date, so we cannot say which figure is newer. The price column also conflicts with its own footnote. The footnote says the figure is for one minute of 1080p video, but the FLUX 3 Max row shows $2.55, which is the 720p promotional rate times 60.
What others say
fal’s landing page, which says it was updated on October 18, carries the heading “Tied for #1 in image-to-video quality.” The release’s headline uses “#1-Tier,” and its body refers to “overlapping confidence intervals among the highest-ranked models.” The model page says “joint #1” and notes that the result covers image-to-video, not text-to-video. Artificial Analysis lists the model second on point estimate. fal calls its results “preliminary,” a word the board does not use.
Unite.AI, bylined to an AI research agent, repeats the release’s figures and notes that the product page reports a different speed figure. It also relays the product page’s “tied for first place” claim, along with the page’s note that MiniMax H3 is listed first on that board, without drawing a conclusion. We found no measurement of its own in the piece. On Hacker News, the one submission we could find, posted October 19, had 2 points and no comments when we checked on October 21.
Our take
For a team that buys video by the second, the budget number is the list price, not the promotional rate. A default five-second 720p clip costs about $0.21 at the promotional rate and $0.85 at list. Against fal’s own Veo 3.1 price of $0.20 per second without audio, the promotional rate is about 79% cheaper, but the list rate is only 15% cheaper. Against Veo 3.1 with audio at $0.40 per second, the list rate is 57.5% cheaper. These comparisons are our arithmetic on fal’s price pages, and those pages do not say whether audio changes FLUX 3 Max’s rate.
The ranking is less settled than the press release suggests. Overlapping intervals mean the two scores are close, not that they tie for first. The board tests image-to-video with audio, not the text-to-video mode the model page leads with, and it pits FLUX 3 Max at 720p against MiniMax H3 at 768p. A new board version is coming, so the places may shift. Calling the model “top-ranked” in product copy claims more than the evidence supports.
The speed figures agree more closely: roughly 11 to 12 seconds for a 10-second clip. Every one of them comes from fal’s or BFL’s own tests, and no independent rerun exists yet. The “22×” figure depends on which timing you take. The model page calls 1080p “optional 1080p upscaling,” which reads as upscaled output rather than native 1080p generation, so teams that need native 1080p should test it themselves.
Who should care? If you produce short clips, ad variants or storyboard previews in volume, the per-second price and the generation time are the numbers you will feel, and both are worth a test. If you pick models by leaderboard rank alone, the current evidence does not justify switching from your default.
How to try it
The shortest route is the text-to-video page for FLUX 3 Max on fal.ai. Keep the default 720p and five-second settings, generate one clip and check what it cost. Audio is on by default, and the page does not say whether turning it off changes the price, so test one clip before any batch run. The same page covers image-to-video and keyframes-to-video.
What to watch
The promotion ends on November 7, after which the list rate applies. Artificial Analysis says its image-to-video board is moving to v2.1 with a new prompt set. We found no independent speed test yet. Two checks would settle the open questions: a rerun of the 10-second timing, and a clear answer on whether audio changes the per-second price.