Start with the arithmetic
Most launch multipliers shrink when you divide the numbers underneath them. This one grows. fal's release pairs 10.95 seconds for FLUX 3 Max with 277.80 seconds for Seedance 2.5, for a 10-second clip. Dividing fal's published figures produces a ratio of about 25.4 to one — greater than the company's "over 22×" headline, which uses a conservative multiplier relative to the two figures shown.
The caution ends there. Both timings come from fal's own latency benchmarks. The release does not describe the hardware, the resolution, the settings, or whether queue time sits inside either figure, and the ratio says nothing about equivalent resolutions, settings, hardware, queue conditions or output quality. A speed comparison published by the company selling one of the models, with no method section, is a claim to test once the API is in your hands. It is not a settled result, however the division works out.
The quality rating is preliminary, and the release says so
fal reports that FLUX 3 Max scored 1,139 Elo for image-to-video in a preliminary evaluation it attributes to Artificial Analysis Video Arena, and says the model achieved that rating across 4,868 votes. It also says confidence intervals overlap among the highest-ranked models. Read plainly, that overlap establishes neither a clear quality advantage for Max nor equivalence with the models around it — and alfw.ai could not independently confirm the 1,139 rating on Artificial Analysis's own leaderboard at the time of writing. Credit where due: the release prints the caveats in the same paragraphs as the claims, which makes the checking easier than usual.
One more separation matters. The latency figures and the Elo come from separate reported evaluations. Without confirmation that they used equivalent output settings and the same model configuration, the published figures do not establish that the reported quality was achieved at the reported 10.95-second latency.
A post-trained model is the actual product here
FLUX 3 Max is not a new foundation model. Black Forest Labs trained FLUX 3. fal Research then post-trained it, with reinforcement learning and inference work, in collaboration with the lab. Against fal's published figures for the base model, Max shows a 50-point higher preliminary Elo and about 11.5-fold lower reported latency. The release does not isolate how much of that latency difference comes from model changes, inference optimisation or serving conditions, so the speedup cannot be credited to post-training alone.
That is the second time fal has run this play. The release frames Max as an extension of the work behind H3 Max, its post-trained version of MiniMax H3, which fal benchmarked at 3.49 seconds — fal's own figure, for a different model under its own conditions. A pattern is forming in video generation: the base model is becoming a starting point, and the version developers actually call is increasingly someone else's optimized variant of it. Pricing, availability and update cadence then belong to the variant's owner, not the lab that trained the base. Access through fal's API also does not, by itself, make Max downloadable or self-hostable, or place it under the base model's licence terms.
What the launch leaves blank
Three blanks matter for anyone deciding today. First, price. The offer is 75% off for the first three weeks from 19 October 2026, and the release never states the price being discounted. A percentage without a base is a mood, not a rate card. Second, specification. The release lists the modes — text, image and keyframes to video — and says nothing about Max's resolution, duration or audio. The base FLUX 3 listing on fal promises up to 20 seconds at 720p or 1080p with native audio, but whether Max matches that is unstated, and the base model's specifications must not be assumed to apply to Max unchanged. We checked fal's endpoint documentation on 19 October 2026 for Max-specific specifications and a live price; neither was available in it. Third, versioning. fal calls this the first iteration and says post-training continues. Before building on the endpoint, confirm whether future revisions will arrive under versioned endpoints or change the model served under the existing identifier — the release does not say.
None of that erases a 10.95-second render, if it holds outside fal's own test. At that speed, video generation stops behaving like a render queue and starts behaving like an interactive tool, where a creator can reject a take and retry inside the same thought. The next useful evidence is independent timing on public settings, a published price, and an Artificial Analysis board entry that no longer needs the word preliminary.
Sources
- [1] fal (primary), "fal Launches FLUX 3 Max, Delivering #1-Tier Image-to-Video Quality at Industry-Leading Speeds", launch release via PR Newswire, 19 October 2026 — post-training method, 1,139 Elo / 4,868 votes / preliminary wording, 10.95s vs 277.80s (Seedance 2.5) and 126.44s (original FLUX 3) latency figures, availability from 19 October 2026, 75% off for the first three weeks, first-iteration status, H3 Max precedent at 3.49sRead source
- [2] fal, "FLUX 3: Video with Native Audio by Black Forest Labs", official model page, checked 19 October 2026 — base FLUX 3 specifications used for context only (up to 20 seconds, 720p / 1080p at 24 FPS, native audio, five endpoints, draft workflow); the FLUX 3 Max release does not state that Max shares these limitsRead source
- [3] fal, model search and endpoint documentation, checked 19 October 2026 — no FLUX 3 Max-specific endpoint documentation with resolution, duration or audio specifications, and no live Max price, was available in it at the time of checkingRead source