FLUX 3 Max and the 22× claim: how to read fal's speed and quality numbers

FLUX 3 Max speed claims deserve a closer look before you change your video workflow. fal reports a large latency advantage over Seedance 2.5, alongside a preliminary arena result near the top group. Those are different measurements: one concerns elapsed generation time on fal, the other concerns human preferences across evaluated outputs.

As of October 21, 2026, the practical question is whether that evidence helps you deliver an acceptable shot sooner. For OmniArt users, start with the available models, their controls and their credit costs. FLUX 3 Max is not in OmniArt's public catalog on that date, and this article does not report measured OmniArt generation times.

What FLUX 3 Max is#

According to fal's announcement (Opens in a new tab), published October 19, 2026, FLUX 3 Max is fal Research's post-trained version of Black Forest Labs' FLUX 3, developed in collaboration with BFL. The announcement is a vendor source, including when it describes benchmark results.

As of October 21, 2026, fal's landing page lists two endpoints at launch, text-to-video and image-to-video ("updated as of October 18, 2026"), and the image-to-video page also mentions a keyframes-to-video option (vendor-reported). The same model page (Opens in a new tab) lists clips from 5–20 seconds; default 720p output and optional 1080p upscaling. The image-to-video endpoint accepts an optional end frame, has audio enabled by default and offers low or high reasoning (vendor-reported). These controls describe fal's product. They do not establish which settings fal used for every timing below.

The three speed numbers, side by side#

All timings in this table are fal's own benchmark figures, checked October 21, 2026. The comparison concerns a 10-second output; the landing page specifically describes an average for 16:9 video at 720p/768p. The final column is our arithmetic, not a separate benchmark.

The landing-page calculation is 277.80 / 12.45 = 22.3; the press-release calculation is 277.80 / 10.95 = 25.4. Thus the advertised 22× multiplier aligns with the landing page after rounding, while the press-release timings imply a larger ratio. Its wording “over 22×” is compatible with that larger ratio. The model page uses an approximate time, so its calculated ratio is approximate too.

The comparison with base FLUX 3 shifts for the same reason. Using fal's reported 126.44-second baseline, our calculations give 126.44 / 12.45 = 10.2 versus 126.44 / 10.95 = 11.5. Do not combine those denominators into a single representative time.

Why the materials differ was not confirmed as of October 21, 2026. Different benchmark snapshots or configurations are possible explanations, but neither is established by these pages. Preserve the sources separately when evaluating the claim.

What fal's 22× claim measures#

The landing-page benchmark (Opens in a new tab), checked October 21, 2026, measures average generation time for its stated configuration on fal's infrastructure (vendor-reported). A clip's output duration is distinct from its generation latency: “10-second video” describes the finished asset, not how long you wait.

The model page explicitly cautions that the benchmark does not guarantee completion time for every request. Resolution, output duration, reasoning settings and operating conditions matter. Check whether a measurement includes queueing, processing and delivery before comparing it with your own stopwatch; the published table does not provide a complete breakdown of those stages.

For a working team, time to an acceptable deliverable also includes retries and review. A quick result that misses a product detail can require another attempt. Write down your acceptance criteria, hold inputs and settings steady, and record the complete wait as well as whether each output passes. This is a suggested evaluation method; we have not run that timing study on OmniArt.

The quality claim: a preliminary arena result#

fal's October 19 announcement reports a preliminary rating of 1,139 Elo across 4,868 votes, with overlapping confidence intervals among the leading models (vendor-reported). That is a dated snapshot of an external evaluation, rather than a permanent ranking.

The Artificial Analysis AA-Video-I2V v1.1 board with audio (Opens in a new tab), checked October 21, 2026, displays this newer snapshot (third-party benchmark):

The displayed intervals and rank ranges overlap, so this snapshot does not clearly separate the top group. That does not prove identical quality or tell you which one handles your shot better. Elo summarizes preferences over the evaluated inputs; it does not measure latency or guarantee subject consistency, usable audio or instruction following on your brief.

Our guide to arena leaderboard claims explains why the evaluation, date and uncertainty belong beside a headline score. Here, the launch's preliminary score and the newer board can both be accurately reported when their dates remain attached.

FLUX 3 Max price: promo versus list#

According to fal's image-to-video page (Opens in a new tab), checked October 21, 2026, the advertised promotion runs October 15–November 7, 2026. The same page gives these subsequent list rates (vendor-reported, USD per output second):

Use the dated promotion for an immediate trial budget and the list rate when planning ongoing work. Recheck the vendor page before spending. Artificial Analysis's API price column uses a separate pricing convention; neither that column nor fal's dollar rates should be treated as OmniArt credits.

A note on H3 Max naming#

As of October 21, 2026, “H3 Max” refers to different products in these sources. fal describes its H3 Max as a fal Research post-trained model; OmniArt's MiniMax H3 Max is MiniMax's own model. We have not confirmed that the Artificial Analysis H3 Max row represents OmniArt's version. Do not transfer fal's timing claims or that row's rating to OmniArt's model based on the shared name.

What this means if you work on OmniArt#

Choose by the deliverable you need to produce. OmniArt's public catalog and pricing configuration, checked October 21, 2026, list these options; the linked model pages are the reader-facing starting points:

These are output rates, not complete quotes for every reference pack. Seedance requests with reference video use 6/13/31 credits per second at 480p/720p/1080p for both output and billable reference-video duration, with at least 4 referenced seconds billed. H3 can charge for referenced video and additional images; H3 Max can charge for reference inputs above its included allowance. Check the workspace's estimate with your actual inputs. All three models require Creator access or above as of October 21, 2026.

H3 Max is not universally the cheapest: at 768p, base H3's 6 credits per output second is below H3 Max's 8. Resolution, reference charges and the number of attempts matter more than a name suggesting a higher tier. No fal latency value in this article estimates your OmniArt wait.

For a starting-image exercise, open the video workspace, choose Seedance 2.5, upload an image you own and select 10 seconds at 720p. Those settings are an example within the catalog checked October 21, 2026. The prompt below directs motion; it is not a speed demonstration, and the prompt handoff does not upload your image or apply the displayed settings automatically.

Judge whether the object stays recognizable and the movement follows the instruction. Repeat with another available model only after matching the settings you can control. Record your own wait separately from fal's published benchmark.