FLUX 3 Max: The 22× Is 12.45 Seconds, Not 10.95.

The 22× on FLUX 3 Max is a division, and the pages do not divide the same two numbers. VentureBeat on October 19 says the claim is a 10-second clip in 12.45 seconds, against 277.80 seconds for Seedance 2.5. Unite.AI’s account of the release says fal measured that same 10-second generation at 10.95 seconds, and described 10.95 as 22 times faster than Seedance 2.5’s 277.80. I divided both. Only one of them is about 22.

What I can check is whether the writeups that did cite those pages are describing one clock.

FLUX 3 Max’s 22× depends on which second you divide

277.80 divided by 12.45 is 22.31. VentureBeat’s “approximately 22 times” is that division, rounded. FourWeekMBA, also October 19, prints the same pair on fal’s table: 12.45 seconds for FLUX 3 Max, 277.80 for Seedance 2.5. Black Forest Labs’ post, as FourWeekMBA quotes it, says “generate a 10-second video in 12 seconds” and “22× faster than Seedance 2.5.” fal’s own post, quoted there, says a 10-second 720p or 768p video in about 12 seconds. About 12, 12, and 12.45 are three ways of pointing at the slower of the two clocks. They are not 10.95.

277.80 divided by 10.95 is 25.37. That is not 22. Unite says the release described 10.95 seconds as 22 times faster than 277.80, at the same reported Elo. If Unite copied the release correctly, the release’s multiplication does not match the release’s inputs. If Unite misread a 12.45 as 10.95, then the 22× in the other two pieces is the number to keep, and Unite has introduced a second clock. I cannot settle which error it is without the release. I can refuse to paste 10.95 and 22× in the same cell.

Unite’s byline says articles by Jonas Reeve are AI-generated and reviewed by the editorial team. I am citing it as a published account of fal’s materials, not as a stopwatch I held. The 10.95 figure does not appear in the VentureBeat piece or the FourWeekMBA piece I opened. Treat it as Unite’s reading of the release until you have the release.

The comparison against the older FLUX 3 splits the same way. Unite says the release called 10.95 seconds 11.5 times faster than the original FLUX 3’s 126.44 seconds. 126.44 divided by 10.95 is 11.55. That one matches the “11.5 times” label. FourWeekMBA says fal’s page describes the cut from 126.44 seconds to 12.45 as about 10 times faster. 126.44 divided by 12.45 is 10.16. Unite also says the product page lists 12.45 and calls the gain over FLUX 3 roughly 10 times. So the 11.5× is glued to 10.95, and the 10× is glued to 12.45. They are not two descriptions of one benchmark. They are two benchmarks, or one benchmark copied wrong.

Kaan Oktay, via Unite, said the team made inference more than 11 times faster while gaining 50 Elo. “More than 11 times” fits 11.55. It does not fit 10.16. If you are writing the speedup into a procurement note, write the seconds next to the multiple. A multiple without the two times is how 22× and 10× and 11.5× end up in one slide.

The Elo is 1,138 or 1,139, and first place is a confidence interval

VentureBeat checked the Artificial Analysis image-to-video leaderboard on October 19. It lists FLUX 3 Max at 1,138 Elo, matching Dreamina Seedance 2.5, two points behind MiniMax H3. The table it prints:

Model	Elo	95% interval
MiniMax H3 (768p)	1,140	1,124–1,156
FLUX 3 Max (720p)	1,138	1,123–1,153
Dreamina Seedance 2.5	1,138	1,121–1,155

Those intervals overlap. VentureBeat says a decisive quality advantage is hard to establish from the displayed results, and that Artificial Analysis places FLUX 3 Max second by its point estimate. fal’s launch materials, which VentureBeat links at fal.ai/flux-3-max, use a preliminary snapshot with slightly different scores and describe shared first place through overlapping intervals.

FourWeekMBA, reading fal’s page rather than the live board, says the page scores FLUX 3 Max at 1,139 Elo and lists MiniMax H3 (768p) at 1,141, Seedance 2.5 at 1,139, and Gemini Omni Flash 1.1 at 1,134. fal notes, in that account, that the tie for first exists because the confidence interval overlaps the leaders, with MiniMax H3 listed first at 1,141. Unite says fal reported 1,139 across 4,868 votes, and that the leaderboard Unite checked listed 1,138 plus or minus 15 across 5,054 samples, second behind MiniMax H3 at 768p with 1,140. The same Unite reading lists Seedance 2.5 at 1,138 and the original FLUX 3 at 1,086.

Count the disagreements before you pick a rank.

The point estimate is 1,138 on the board VentureBeat and Unite describe, and 1,139 on the fal page FourWeekMBA and Unite describe. One point. MiniMax H3 is 1,140 on that board and 1,141 on the product-page table. The original FLUX 3 is 1,086 on the board Unite read and 1,089 in the before-and-after fal is using for the 50-point gain. 1,089 to 1,139 is 50. 1,086 to 1,138 is 52. The “50-point” sentence belongs to fal’s pair, not to the board’s pair.

The vote counts are not the same object either. 4,868 votes is what Unite says fal reported for the 1,139. 5,054 samples is what Unite says the board showed for the 1,138. I am not averaging them. A later board can have more votes and a different point estimate. A launch graphic can freeze an earlier snapshot. Paste the date and the page, or do not paste the Elo.

“Tied for first” is fal’s description of overlapping intervals, not a claim that the point estimate is alone at the top. VentureBeat’s board has MiniMax H3 two points ahead. FourWeekMBA’s reading of fal’s table has MiniMax H3 two points ahead, at 1,141 against 1,139, and still calls the overlap a tie. Both can be printed. Neither is “we won outright.” If your caption says number one, you have upgraded a tie-by-interval into a rank the point estimate did not give you.

The resolutions are not identical either. VentureBeat says fal’s timing comparison is an average for 10-second 16:9 clips at 720p or 768p, depending on the model. The Elo table labels MiniMax H3 at 768p and FLUX 3 Max at 720p. Seedance 2.5’s row in VentureBeat’s table has no resolution in the cell I copied. “Same quality” in the Black Forest Labs post is an Elo claim. It is not a claim that every clip was the same pixel count. I would not write “same settings” unless the page you opened says the settings matched.

The discount is 19 days, or two weeks, or three

The price that survives across the three pieces is $0.0425 per second of 720p output, against a $0.17 list price, 75 percent off, with November 7 on the page. VentureBeat’s table makes the 10-second arithmetic explicit: promotional $0.425, regular $1.70. I multiplied $0.0425 by 10. That is $0.425, not $0.43. FourWeekMBA says a 10-second clip costs about $0.43 at launch. “About $0.43” is a rounding of $0.425. If you are putting the rate into a spreadsheet, use $0.425 or $0.0425 times seconds. Do not use the rounded sentence as a rate card.

VentureBeat also prints the upscale. 1080p, upscaled, is $0.0725 promotional and $0.29 regular, so a 10-second upscale is $0.725 now and $2.90 after November 7. A note that says “under fifty cents for 10 seconds” is the 720p promotional row only. It is not the upscale, and it is not the list price.

The calendar does not match the slogan. fal’s post, via FourWeekMBA, says 75 percent off for the next three weeks. Black Forest Labs, via the same piece, says 75 percent off for the next two weeks. The page both of them are pointing at ends the rate on November 7, which is 19 days after the October 19 launch. Three weeks is 21 days. Two weeks is 14. Nineteen is neither. Unite says the product page runs until November 7, 2026, and the release describes the discount as the first three weeks. This site has already watched a public price with an end date get treated as a permanent card. Argon’s $2 ended on a printed day, not on a vibe. November 7 is the day in the rate line. “Two weeks” and “three weeks” are posts. If you are signing a usage estimate, use the date.

FourWeekMBA’s 79 percent headline is its own arithmetic: $0.0425 against Veo 3.1 at $0.20 per second without audio, on fal’s pages. $0.0425 is 21.25 percent of $0.20, so the launch rate is 78.75 percent below that Veo rate. “About 79 percent” is that gap, rounded. It is not 79 percent below every video rate on the site, and it is not the list price. $0.17 against $0.20 is 15 percent below, not 79. After November 7, the headline comparison collapses unless you recompute it.

The Veo line FourWeekMBA quotes is “$0.20 without audio or $0.40 with audio for 720p or 1080p.” The 79 percent uses the without-audio number, and that number covers 720p or 1080p, while FLUX 3 Max’s $0.0425 is 720p. FourWeekMBA says so. A 1080p FLUX clip on VentureBeat’s table is $0.0725 promotional, not $0.0425. If you need 1080p, you are not in the cell the 79 percent was built from.

Other rows on fal’s pricing table, as FourWeekMBA read it on October 19, are cheaper per second than the launch rate or the list rate. Kling Video v3 Pro is $0.14. MiniMax H3 Max is $0.05. H3 Max Turbo is $0.025. fal’s post called the discounted model the most affordable frontier model at its quality level. FourWeekMBA’s reading is that H3 Max Turbo is cheaper per second than the launch rate, and that the $0.17 list price costs more per second than Kling’s $0.14. “Most affordable” is a sentence. The table is a set of rates. They are not the same object. Veo 3.1 Fast is listed from $0.10 per second, which is also above $0.0425 and below $0.17. A from-price is not a quote. I would not budget against “from $0.10” without the row that applies to the clip you actually want.

Ten versions of a 10-second 720p shot, at VentureBeat’s figures, cost $4.25 during the promotion and $17 after. That excludes staff time. VentureBeat’s point, which I agree with as a budgeting warning and not as a quality claim, is that a faster model can raise spend if you generate more variants and accept the same share. The 22× does not cap the invoice.

Two endpoints, or three, and a reasoning dial the multiple ignores

Unite says the announcement describes text-to-video, image-to-video, and keyframes-to-video, while the product page lists two endpoints at launch, text to video and image to video, both at 720p. VentureBeat describes public endpoints for text-to-video and image-to-video, and then a separate keyframes endpoint that accepts up to 10 images. The writeups disagree on whether keyframes is a launched endpoint or an announced one. If your integration plan assumes three routes on day one, you are following VentureBeat’s reading and the announcement, not Unite’s reading of the product page.

Output length, where the pieces agree, is five to 20 seconds. Generation is 720p, with optional upscaling to 1080p. Audio can be on or off. Formats include landscape, portrait, and square. Image-to-video can take an ending image, in VentureBeat’s account. None of that is a speed. A 20-second clip is not a 10-second clip with the 12.45 reused. Do not scale 12.45 by two and call it measured.

VentureBeat also says the text-to-video documentation has low and high reasoning settings, and that the higher setting increases generation time. The 12.45 and the 10.95 are not labeled with a reasoning setting in the sentences I have. A multiple that ignores a dial the vendor documented is a multiple for one setting, whichever setting the table used. I cannot tell you which. I can tell you not to quote 12.45 as the time for the high setting unless the page you are quoting says that.

API commercial use is not a model license. VentureBeat says fal allows commercial use of API-generated content under its terms, and that this is not a license to download, modify, or operate the model independently. If your counsel needs the weights, the launch posts I am using do not grant them. A sticker price and a batch price were already different objects on this site. A generation right and a weight right are the same kind of split.

What to put in the cell

If the cell is speed against Seedance 2.5, use 12.45 and 277.80, and say the 22× is 277.80 divided by 12.45, approximately. Do not put 10.95 in that cell unless you have opened the release and confirmed Unite’s reading. If you do confirm 10.95, the multiple is about 25, not 22, and you should print 25.

If the cell is rank, print the point estimate and the interval, and say MiniMax H3 is ahead on the point estimate in every table I saw. “Tied for first” means the intervals overlap. It does not mean the point estimates match.

If the cell is price, use $0.0425 per second of 720p until November 7, then $0.17. Use $0.0725 and $0.29 if you are upscaling to 1080p. Do not import the 79 percent unless you are comparing the 720p launch rate to Veo’s $0.20 without-audio rate, and say so. Do not import “two weeks” or “three weeks” over the date.

No clip was generated for this article. I have not voted in the arena. The useful fact this week is that the multiple, the Elo, and the discount length each exist in more than one version, and the versions are on pages published the same day.