The fal Research team has post-trained the FLUX 3 video model from Black Forest Labs and published the result under the name FLUX 3 Max. According to fal's own measurements, the model assembles a 10-second 720p clip with sound in 12.45 seconds — approximately 22 times faster than Seedance 2.5 in the same test. The quality of the upgrade has been independently confirmed by the Artificial Analysis arena, where FLUX 3 Max landed in the leader group. Until November 7, 2026, generation is sold with a 75% discount, making such a clip cost $0.425.

What happened

fal Research refined the already released FLUX 3 video model through post-training in collaboration with Black Forest Labs and opened it via the fal API under the name FLUX 3 Max. The model creates videos with sound in two modes: text-to-video and image-to-video. After inference optimization, latency on a 10-second 720p fragment decreased from 126.44 to 12.45 seconds, a reduction of approximately tenfold. According to Artificial Analysis arena data, post-training raised the Elo of the base FLUX 3 from 1089 to approximately 1139. The developer also published pricing: $0.0425 per second of 720p generation under the promo until November 7, 2026, and $0.17 per second after the promotion.

Context

The main independent benchmark for quality here remains the Artificial Analysis leaderboard for image-to-video (AA-Video-I2V v1.1). On it, FLUX 3 Max scored an Elo of 1138 and was in the same group as Seedance 2.5 (1138) and MiniMax H3 (1140): a spread of one to two points does not allow naming a single leader. Notably, this result was achieved through post-training on top of an already released base model, rather than a new architecture: post-training has recently remained the cheapest way to catch up with competitors after a release. Latency is not separately tracked in such rating tables, so speed figures currently exist only in fal's materials.

Why this matters for the industry

For the industry, the intersection of the interactivity threshold is more important than the quality increase: the "generated, watched, corrected" cycle now fits within a live user session and an agent's working cycle, which is the foundation for batches of dozens of scene variants with automatic selection. If fal's measurements are reproducible, the bottleneck in pipelines will become queues and provider throughput, not the model itself. Competition among video models at equal quality shifts toward speed and price, making a counter-move from Seedance 2.5 and MiniMax H3 logical: accelerating inference and lowering rates. A wave of post-trainings on top of already released base video models is also likely, as well as the appearance of latency measurements alongside Elo in third-party ratings.

Why this matters for users

The model can be tested right now on fal.ai, where text-to-video and image-to-video modes are open, and a 10-second clip is assembled in approximately twelve seconds. At the promo price, such a clip costs $0.425, and after November 7, 2026, it will be $1.70. This makes fast scene iterations, portrait animation, and product videography cheap: a hypothesis becomes a visual prototype in minutes and dollars. A reasonable order of actions: first complete a trial session at the promo price, then measure latency on your own prompts and load, and only then calculate unit economics based on the post-promotion price.

What is currently unknown / limitations

All speed figures, including the 12.45-second latency, the previous 126.44 seconds, and the "22 times faster than Seedance 2.5" multiplier, were obtained by fal itself: the published materials do not include hardware, prompt sets, number of runs, or competitor launch settings, so until independent replication, these are vendor claims. The one-to-two point Elo gap with Seedance 2.5 and MiniMax H3 is statistically indistinguishable; it is correct to speak only of entering the leader group, not of superiority. Sound quality has not been evaluated separately anywhere: the arena covers the visual part of image-to-video, and there is no complete audio evaluation methodology in the sources. Finally, the promo rate is valid only until November 7, 2026, after which the cost of generation will increase fourfold.

Sources

Author

Look at AI, editorial team