On October 19, 2026, fal Research and Black Forest Labs released FLUX 3 Max. The headline is speed: a 10-second 16:9 video comes back in roughly 12 seconds, against 277.80 seconds for Seedance 2.5 in the same official table, an official gap of about 22x.
But fal's and BFL's own pages carry four speed figures, not one: 12.45 seconds on the fal landing page, 10.95 seconds in the fal newsroom, approximately 11 seconds on the fal API model page, 12 seconds on the Black Forest Labs site. Pricing is no cleaner: fal's pricing table row reads | FLUX 3 Max | seconds | $0.085 | $0.425 | 11.76 |, matching neither the $0.0425 promo rate nor the $0.17 list price, and fal does not explain it.
More important, 277.80 seconds is fal's unilateral number: ByteDance has never published a Seedance 2.5 generation-time benchmark, and fal never says which resolution tier, aspect ratio, prompt length, or audio state it covers. The "same quality" premise was never closed.
This review does not declare a winner. It pins every number to its source: how many official sets exist, which fields align, which never can, and what the bill looks like. On scope, this site has covered the 30-Second Club, a selection piece, and Open Video Model Local Deployment, a licensing and VRAM piece; this article is the speed and cost angle. See also six-way video API cost comparison and the Seedance 2.5 vs MiniMax H3 review.
1. What it actually is: a post-trained model
FLUX 3 Max is not a new foundation model. Black Forest Labs provides the FLUX 3 base model; fal Research does the post-training, reinforcement learning and inference optimization on top.
BFL CEO Robin Rombach: "We built FLUX 3 as a foundation for others to build on." FLUX 3 Max, he said, is that in practice, with fal customizing the model for its own platform. fal CEO Burkay Gur added that the biggest advances in generative AI may come from "optimizing models and the systems that run them together." fal research lead Kaan Oktay was blunter: model performance does not stop improving when base model training ends. The conclusion: FLUX 3 Max sells efficiency from post-training plus inference optimization, not a leap in parameter count or architecture. This is the first iteration, post-training continues, and every number here may still move.
2. The full official speed and Elo table
The fal model page publishes a six-model table: Elo from Artificial Analysis's AA-Video-I2V v1.1, preliminary, times averaged for a 10-second 16:9 video at 720p/768p.
| Model | Image-to-video Elo | Time for 10s video |
|---|---|---|
| FLUX 3 Max | 1,139 | 12.45s |
| Gemini Omni Flash 1.1 | 1,134 | 47.37s |
| MiniMax H3 (768p) | 1,141 | 83.26s |
| FLUX 3 | 1,089 | 126.44s |
| Wan 3.0 | 1,127 | 180.16s |
| Seedance 2.5 | 1,139 | 277.80s |
First, the official page footnotes one model faster than everything listed: H3 Max, another post-trained fal Research model, at 7.55 seconds. FLUX 3 Max is fastest only inside this table, and the table-wide fastest entry is a sibling from the same lab.
Second, the leaderboard's number one is not FLUX 3 Max. It ranks MiniMax H3 first at Elo 1,141, with FLUX 3 Max on 1,139, tied with Seedance 2.5, and notes the confidence interval overlaps the leader's. The official claim is "tied for first," which the overlap makes fair, but the ranking does put MiniMax H3 on top; both are true and must be stated separately. The Elo carries 4,868 votes.
Third, both sit at Elo 1,139 while their times differ by the official 22x ratio, so speed is the only dimension this table separates.
3. The official sources give four speed numbers
The problem is not outside scrutiny of fal, but fal's own four pages carrying four answers.
| Source | Time for 10s video | Official multiplier |
|---|---|---|
| fal landing page | 12.45 seconds | about 22x |
| fal newsroom | 10.95 seconds | over 22× |
| fal API model page | approximately 11 seconds | more than 25× |
| Black Forest Labs site | 12 seconds | 22× |
Four numbers for one operation, multipliers spanning 10x to 25x, no shared timing convention. The other half of the disagreement is the comparison against the original FLUX 3: the model page calls 126.44 seconds down to 12.45 seconds "about 10x," the newsroom calls the same 126.44 seconds down to 10.95 seconds "11.5x." The denominator is fixed, so the two numerators land on 10x and 11.5x. The Seedance 2.5 multiplier splits the same way: most official pages say 22x, the fal API model page writes more than 25×.
fal ships a disclaimer to read alongside all of it: "The latency comparison describes the tested benchmark configuration, rather than a guaranteed completion time for every request." Generation time varies with request settings, and fal never reconciles the four sets.
A third-party measurement offers a third answer: the overchat write-up reports being about 10x faster than Seedance 2.5 and about 4x faster than FLUX 3, and observed that Seedance 2.5 follows prompts more tightly and looks flashier. With 22x, 25x, 11.5x and a third-party 10x in play, the honest move is to attribute each.
4. The key gap: 277.80 seconds is a one-sided number
Every multiplier above rests on one premise, that 12.45 and 277.80 seconds describe the same quality. That premise does not hold. The 277.80 seconds is fal's own entry in its own table; ByteDance has never published a Seedance 2.5 generation-time benchmark, no public third party has reproduced it, and fal does not say which resolution tier, aspect ratio, prompt length or audio state it covers. Not one of those variables is locked.
So the table supports this: under fal's own test configuration, FLUX 3 Max waits substantially less than Seedance 2.5. It does not support the stronger claim of being 22x faster at the same quality, which would need a ByteDance baseline that does not exist. The third-party 10x sits closer to felt experience, so the strength of the 22x depends on how you measure.
5. Mechanism differences: reasoning_level and timings.inference
A subtler comparability gap sits inside the timing convention. FLUX 3 Max has a reasoning_level parameter: the default is high, it is slower, and it includes prompt expansion. Seedance 2.5 has no equivalent, and fal has not published whether the 12.45 seconds includes expansion time. fal's docs define timings.inference as the seconds spent denoising the video, explicitly excluding prompt expansion, decoding and upload. So the official timing field does not cover expansion or upload in the first place, and whether the default high setting adds wall-clock time is unpublished. Comparing 12.45 against 277.80 therefore measures two different inference paths, and neither side publishes expansion or post-processing overhead: the multiplier is an order-of-magnitude reference, not a performance metric.
6. What can be aligned and what cannot
| Alignable field | FLUX 3 Max | Seedance 2.5 |
|---|---|---|
| Image-to-video Elo (same table, same method) | 1,139 | 1,139 |
| Time for a 10s video | 12.45s | 277.80s |
| Duration cap | 5-20 seconds | 4-30 seconds |
| Frame rate | 24 | 24 |
| Native audio | on by default | on by default |
| 1080p nature | upscale, not native render | not published by the vendor |
The dimensions that do not line up outnumber those that do. Reference input is not comparable: FLUX 3 Max has no reference-to-video endpoint and no published cap on reference material, while Seedance 2.5 accepts up to 50 references, 30 images, 10 videos and 10 audio files. Keyframe mechanisms differ too: Max uses a standalone keyframes endpoint of up to 10 frames at 24fps addressed by frame_index, Seedance 2.5 uses reference-to-video plus prompt timestamp choreography. Max also has no video editing, continuation or chroma key endpoints, while Seedance 2.5 does. Hardware, API tier, queue time, audio toggle, prompt length and sample size are unpublished on both sides, and Max's 1080p is an upscale, not the base model's native 4K UHD tier.
7. The money: four official unit prices
Pricing comes in four numbers from the fal model page, checked October 22, 2026.
| Tier | Promotional | List |
|---|---|---|
| 720p | $0.0425 / second | $0.17 / second |
| 1080p | $0.0725 / second | $0.29 / second |
The discount window runs from October 15 to November 7, described as 75% off for a limited time. Billing follows output duration, with no minimum spend and no subscription, and text-to-video and image-to-video are priced the same. Converted from the official per-second rate, a 10-second video costs $0.425 at the 720p promo rate, $1.70 at 720p list, $0.725 at 1080p promo, and $2.90 at 1080p list.
Neither common video API billing pattern, a monthly seconds bundle or a per-call purchase with a minimum, applies here. Trial cost is a fraction of a dollar per ten seconds at 720p; at scale, unit price and concurrency limits decide margin.
One row has to be quoted as-is: fal.ai/pricing lists FLUX 3 Max as | FLUX 3 Max | seconds | $0.085 | $0.425 | 11.76 |, matching neither model-page rate, with no explanation of what it counts. The official unit price therefore exists in three versions on the same site, so a quote should name its source page.
8. What is still undisclosed
The full resolution tier list is not disclosed, since official pricing covers 720p and 1080p only. Neither is the official duration range: model page examples show 6 seconds and 10 seconds but no range. Native audio, and concurrency and queue limits, are also undisclosed. On reference caps, the keyframes endpoint states a 10-frame cap, but no Max-side cap is published.
The media line, unconfirmed by the official side, adds landscape and portrait aspect ratios, an audio toggle, a 5-20 second duration range, native audio and a maximum of 10 keyframes. What media reported as "1080p $0.29 per second" is now officially confirmed as the 1080p list price rather than a promo rate, which matters when quoting it.
9. Capability, integration, and channels
The official unified API on fal supports text-to-video, image-to-video and keyframes-to-video, with the endpoint blackforestlabs/flux-3-max/text-to-video called through fal.subscribe, and API output is licensed for commercial use under the service terms.
One line needs to be drawn carefully. The FLUX 3 base model has its own feature set: draft mode for quick previews, multi-scene generation, video editing, agentic chaining, video continuation, and output up to 4K (UHD). These belong to the base model, not to FLUX 3 Max: Max has only the 720p and 1080p price points, and 4K appears on no official Max page.
The channel reality matters too. BFL publishes no dedicated price list for FLUX 3 Max and no matching API endpoint, and the public endpoints on docs.bfl.ai are flux-3-video, flux-3-image and flux-3-image-klein. BFL's own FLUX 3 Video pricing has two official sets, listed side by side without arbitration. FLUX 3 Max has no official blog post and no ComfyUI official Partner Nodes entry, and FLUX 3 weights there still read "Open weights are coming soon." fal is the only route to Max.
10. Quick selection guide
| Your situation | Lean toward | Basis |
|---|---|---|
| Per-second billing, batch generation | FLUX 3 Max | 720p at $0.0425 per second promo, $0.425 for 10s, no minimum spend |
| Interactive previews, long waits unacceptable | FLUX 3 Max | fal table: 12.45s against 277.80s, official 22x (also 25x, third-party 10x) |
| Prompt adherence and final look first | Seedance 2.5 | Third-party overchat observation: tighter prompt following (not official) |
| Need reference inputs or keyframe choreography | Seedance 2.5 | Up to 50 references, 4-30 seconds; Max has no reference-to-video endpoint |
| Need 1080p | Either | 1080p at $0.0725 per second promo, but Max upscales 1080p |
| Want self-hosting | Out of scope | Max is closed source, fal-only; BFL ships no endpoint or price list |
FAQ
Q1: How many official speed numbers are there? A1: Four, as section 3 lists them: the fal landing page, the fal newsroom, the fal API model page and the BFL site. Their multipliers span 10x to 25x with no official reconciliation, so they can only be quoted side by side.
Q2: Who is number one on the leaderboard? A2: The leaderboard puts MiniMax H3 first at Elo 1,141, with FLUX 3 Max at 1,139, tied with Seedance 2.5, and notes the confidence interval overlaps the leader's. The official "tied for first" claim and the ranking that puts MiniMax H3 on top therefore both hold and must be stated separately. It is the preliminary AA-Video-I2V v1.1 result from 4,868 votes.
Q3: Is the 277.80-second comparison usable?
A3: As an order-of-magnitude reference, yes. It is fal's one-sided number; ByteDance has never published a Seedance 2.5 generation-time benchmark, and fal does not say which resolution tier, aspect ratio, prompt length or audio state it covers. Seedance 2.5 has no reasoning_level parameter, and fal's timings.inference excludes prompt expansion, decoding and upload.
Q4: What does a 10-second video cost? A4: Converted from the official per-second rate, 720p promotional at $0.0425 gives $0.425, 720p list at $0.17 gives $1.70, 1080p promotional at $0.0725 gives $0.725, and 1080p list at $0.29 gives $2.90, with the discount window running October 15 to November 7 at 75% off. Billing follows output duration, with no minimum spend and no subscription. And fal's pricing row of $0.085 / $0.425 / 11.76 matches neither model-page tier, so a quote should name its source page.
Q5: Can FLUX 3 Max be used commercially, and what is still missing? A5: API output is licensed for commercial use under the service terms, and the model is closed source and reachable only through fal, since BFL ships no dedicated price list and no API endpoint for it. The official side has not disclosed the full resolution tier list, the official duration range, native audio, or concurrency and queue limits. BFL's own FLUX 3 Video pricing has two official sets.