fal and Germany-based AI lab Black Forest Labs (BFL), the team behind the FLUX model family, have launched FLUX 3 Max, a video generation model that fal says can create a 10-second clip in about 12 seconds. The model is being offered at a discounted rate through Nov. 7, with 720p priced at $0.0425 per second during the promotion.
Still, the two main claims attached to the release both come with limits. The statement that FLUX 3 Max is tied for the best image quality applies only to the image-to-video leaderboard, and even there it is a statistical tie. The claim that it is 22 times faster than ByteDance’s Seedance 2.5 comes from fal’s own benchmark, while another chart posted later by fal showed a narrower gap of about 5.5 times under what it described as the same conditions.
FLUX 3 Max is a post-trained version built on FLUX 3
BFL opened access to its FLUX 3 video model on Aug. 4. According to fal’s launch page, fal Research worked with BFL to apply post-training on top of FLUX 3. In practical terms, that means continuing training on an existing model with new data and new objectives, rather than building a model from scratch.
fal described three layers of work. The first was post-training aimed at improving visual detail, motion consistency, and image-to-video quality. The second was inference optimization, which it said reduced the time needed to generate a 10-second video from 126.44 seconds on the original FLUX 3 to 12.45 seconds. The third was system-level tuning on fal’s own inference engine.
On the product side, fal’s API documentation lists three endpoints: text-to-video, image-to-video, and keyframe-to-video. The keyframe mode supports up to 10 images to define frames at different points in a clip. Supported video length ranges from 5 to 20 seconds at 720p, with an option to upscale output to 1080p.
Pricing is based on output seconds, with no minimum spend and no subscription requirement. Text-to-video and image-to-video are priced the same. From Oct. 15 to Nov. 7, fal is offering a 75% discount. That puts 720p at $0.0425 per second, down from $0.17, and 1080p at $0.0725 per second, down from $0.29. Commercial use is allowed, but the model weights are not open, so access is limited to fal’s API and web playground.
Quality claim holds only on the image-to-video leaderboard
Independent benchmarking site Artificial Analysis uses blind comparisons in which reviewers choose the better of two videos, then aggregates those matchups into an Elo-style score. Because small score differences can fall within statistical noise, the site also publishes 95% confidence intervals.
On the image-to-video leaderboard, MiniMax H3 (768p) scored 1,140, with a confidence interval of 1,124 to 1,156. FLUX 3 Max (720p) scored 1,138, with a range of 1,123 to 1,153. Seedance 2.5 also scored 1,138, with a range of 1,121 to 1,155. Those intervals overlap heavily, which is the basis for the claim that FLUX 3 Max is effectively tied at the top in that category.
The picture changes on the text-to-video leaderboard. Wan 3.0 ranked first at 1,173, followed by Utopai X at 1,166. Seedance 2.5 and FLUX 3 Max both posted 1,159, with FLUX 3 Max listed in fourth place.
The post-training step did improve the base model’s benchmark results. On the image-to-video leaderboard, the original FLUX 3 scored 1,086, meaning FLUX 3 Max gained 52 points. On the text-to-video leaderboard, the score rose from 1,145 to 1,159, an increase of 14 points.
The 22x speed figure comes from one fal benchmark
The first set of timing numbers appeared on fal’s launch page and in an image attached to its launch post. The footnote said the benchmark covered average generation time for 10-second, 16:9, 720p or 768p videos. In that chart, FLUX 3 Max took 12.45 seconds, MiniMax H3 took 83.26 seconds, the original FLUX 3 took 126.44 seconds, Wan 3.0 took 180.16 seconds, and Seedance 2.5 took 277.80 seconds. Dividing 277.80 by 12.45 gives roughly 22, which is where the headline claim came from.
Less than three hours later, fal posted another speed chart on X with what it described as the same footnote conditions. In that version, FLUX 3 Max still showed 12.45 seconds, but Seedance 2.5 dropped to 69.09 seconds, reducing the gap to about 5.5 times. The original FLUX 3 was listed at 46.29 seconds, or about 3.7 times slower than FLUX 3 Max. MiniMax H3 moved in the other direction, from 83.26 seconds to 122.0 seconds.
Both sets of numbers were self-reported by fal. The company did not explain why the results differed, and it did not specify when the tests were run or where they were executed. For Seedance 2.5 alone, the two official charts differ by about four times. Another missing detail is the inference setting. The model has low and high modes, and the default high mode automatically expands prompts. Higher settings take longer to run, but the launch page did not say which mode was used to reach the 12.45-second figure.
That also means 12.45 seconds is still longer than the 10-second clip being generated. It is close to real time, but not real time. Teams considering adoption would still need to test it against their own workloads.
Discounted pricing undercuts some rivals
At list price, 720p output costs $0.17 per second, the same as BFL’s own FLUX 3 standard HD offering. At the promotional rate of $0.0425 per second, it is lower than BFL’s own draft-mode HD price of $0.06 per second. The source article noted that draft mode is meant for quick, lower-cost previews before rendering a full-quality version, while fal’s pitch is that users do not need separate fast models for drafts and slower models for final output.
On competing products, the Artificial Analysis image-to-video leaderboard lists API pricing of about $0.09 per second for MiniMax H3 and about $0.41 per second for Seedance 2.5. On that basis, FLUX 3 Max is cheaper than both during the promotional period.