Seedance 2.5 vs FLUX 3 Video: 22× Faster Is a Claim, Not a Measurement

Black Forest Labs put a number on the internet in October that does most of the marketing work for FLUX 3 Video: "22× faster than Seedance 2.5 at the same quality." It sits in the launch post for FLUX 3 Max, next to a second claim — a 10-second video generated in 12 seconds — and both are vendor statements about the vendor's own product. Neither has a published methodology, a fixed test set, or a quality criterion anyone outside BFL can reproduce. Seedance 2.5, for its part, has a measured position on three Artificial Analysis boards and a rate card that costs more than twice as much per minute. Which of those two facts should move a production decision is the whole question, and the answer is not the one the multiplier suggests.

The two models were not built for the same call. Seedance 2.5, released by ByteDance on 31 July 2026, is a reference-and-editing model: 30-second clips, 50 multimodal references, audio you can supply as a standalone track. FLUX 3 Video, whose generation capabilities BFL documented on 4 August 2026, is a resolution-ladder model: HD through 4K priced by pixels per frame, native audio at no extra charge, a Draft tier for cheap iteration, and — in the FLUX 3 Max tier that arrived in October — a speed story that no third party has yet tested.

Start with the one board where both actually appear

Artificial Analysis runs three video boards on separate Elo scales, read on 12 November 2026:

• Text-to-video with audio, AA-Video-T2V v2.1 — Seedance 2.5 fourth at Elo 1,155 (±13, interval 1,142–1,168) over 7,765 votes and $34.12 per minute; FLUX 3 Max sixth at 1,153 (±15, 1,138–1,168) over 4,094 votes and $8.70 per minute; FLUX 3 tenth at 1,137 (±13, 1,124–1,150) over 6,827 votes and $17.40 per minute.

• Image-to-video with audio, AA-Video-I2V v2.1 — Seedance 2.5 first at 1,161 (±15, 1,146–1,176) over 6,540 votes; FLUX 3 Max sixteenth at 1,068 (±14, 1,054–1,082) over 6,763 votes; FLUX 3 fifteenth at 1,071 (±14, 1,057–1,085) over 6,757 votes.

• Video editing, AA-Video-Editing v1.1 — Seedance 2.5 fourth at 1,147 (±11, 1,136–1,158) and $40.82 per minute; neither FLUX 3 entry appears on this board at all.

Two things fall out of that. On text-to-video the two flagships are statistically tied — 1,155 against 1,153, intervals almost entirely overlapping — while FLUX 3 Max costs a quarter of what Seedance 2.5 costs per minute. On image-to-video, where BFL's own October post claims FLUX 3 Max is "tied for #1", the board puts it sixteenth, 93 Elo points behind Seedance 2.5 and below a dozen models including HeyGen Video 1 at $2.40 per minute. That is not a small discrepancy between marketing and measurement. It is the largest gap between a vendor claim and an independent reading anywhere in this comparison, and it is the reason to treat the 22× figure as a hypothesis rather than a spec.

The board's absence line matters too: FLUX 3 Video has no editing row. Seedance 2.5's 1,147 there — fourth of eleven, behind Wan 3.0, Gemini Omni Flash 1.2 and Hyperflow H3 — is a measured position on a task the other model is not scored on. Absence is not inability, but a pipeline decision cannot be built on the absence of a denial either.

What "22× faster" would have to mean to be useful

Speed claims in video generation are unusually slippery, because the thing being measured is never stated. Wall-clock time per clip depends on resolution, duration, whether audio is generated, queue depth and whether the render is a draft or a final. A 10-second clip in 12 seconds at 720p is roughly real-time and genuinely fast. The same model at 4K with audio on a busy afternoon is not, and nothing in the FLUX 3 Max post says which configuration the 12 seconds and the 22× refer to.

Even taken at face value, the number has a narrow use. If your workflow generates a hundred candidate shots and a human picks one, generation latency dominates and a 22× advantage is transformative — the loop closes inside a meeting. If your workflow generates four shots for a client review, twelve seconds versus four minutes changes nothing about the day, and the 93-point image-to-video gap changes everything about whether the chosen shot is usable. Which of those two is your business is the actual decision, and no multiplier answers it.

There is a fair version of the speed argument, and BFL does not make it explicitly. FLUX 3 Video's Draft mode — HD previews at $0.06 per second against $0.17 for full HD — is the same shape of cost control as Seedance 2.5's 480p Draft workflow, and it costs less per preview second than Seedance's roughly $0.10. Two models with a cheap-iteration tier and a final-render tier is a mature design; the interesting question is which one's preview is predictive of its final, and neither vendor publishes that.

The input contract is where the models genuinely diverge

Seedance 2.5 accepts up to 50 reference assets in a single request: 30 images, 10 video clips totalling no more than 30 seconds, and 10 audio clips. Reference video is capped at 200 MB per file, input frame rates run 24–60 fps, and untextured 3D clay renders can be supplied as a structural reference that fixes spatial layout, poses, motion paths and camera angle before lighting is applied. Audio references stand alone — you do not need an accompanying image or video to use one.

FLUX 3 Video's documented surface is a different shape. It covers text-to-video, image-to-video, video-to-video, video continuation and keyframe transitions, multi-scene clips and multiple camera angles, with native audio including multilingual dialogue, lip sync and sound effects generated alongside the picture at no extra charge. That is a broader set of task types and a narrower control surface per request — BFL documents continuation and keyframe transitions, not a fifty-asset reference budget with per-asset weighting.

Practically: if your problem is holding a named character, a specific product and a music bed across a long sequence, the reference budget is the deciding column. If your problem is taking an existing clip and continuing it, or authoring keyframe transitions, FLUX 3 documents that and Seedance 2.5 handles it through an extension task rather than a first-class input. Neither is strictly more capable; they are aimed at different stages of a production.

Priced per accepted clip, not per second

At list rates the gap is stark and it does not favour Seedance. BytePlus's worked examples put Seedance 2.5 at $0.103 per second at 480p, $0.231 at 720p and $0.569 at 1080p for 5-second clips with no video input; Artificial Analysis normalises it to $34.12 per minute at 1080p, or $40.82 on editing work. FLUX 3 Video publishes a full resolution ladder: $0.06 per second for Draft HD, $0.17 for HD, $0.29 for FHD, $0.40 for QHD and $0.80 for UHD on text-or-image-to-video, with video-to-video running $0.12 to $0.95. A five-second HD clip is $0.85.

Put them on one 30-second deliverable. Seedance 2.5 makes it in a single pass at 1080p for about $17.07 at the vendor's own per-second rate. FLUX 3 Video makes a 20-second clip at 1080p for $5.80 and needs a second generation to reach 30 seconds — roughly $11.60, at the cost of a continuity seam. At UHD the FLUX rate is $0.80 per second, $24.00 for thirty seconds, which is the only published 4K-class rate in this comparison that a customer can actually be charged.

That last point is where the two models stop being comparable. Seedance 2.5 in 4K went live on Runway and Picsart on 11 November 2026, priced by those platforms, but BytePlus ModelArk's own rate card still ends at 1080p for dreamina-seedance-2-5-260628 and Volcano Engine's film-facing 4K tier remains an invited trial with no published price. FLUX 3 Video publishes a 4K rate, sells it, and does not require an invitation. If 4K delivery is the requirement, this comparison is over before the quality argument starts.

The integration risk nobody puts in the comparison

Two vendor contracts, two authentication models, two job systems, and one shared failure: a generation that dies mid-queue after you have paid for the queue. Seedance 2.5 will not run at all until the BytePlus account carries a balance above $30, an AI Savings Plan at the $30 tier or above, or a purchased resource pack — an API key on its own is not enough, which is worth knowing before a spike. FLUX 3 Video runs off the BFL dashboard and API with no subscription or seat fees, and enterprise terms add volume discounts and SLAs.

Whichever you pick, the decision gets cheaper to revisit if the orchestration around the video model is not welded to it. That layer — prompt expansion, reference captioning, the shot list, the review pass that decides which draft gets the expensive render — is text-model work. OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, so the review step can be a cheap model from one vendor and the prompt work a stronger one from another without a second contract, and automatic failover keeps the loop alive when a provider wobbles. Neither Seedance 2.5 nor FLUX 3 Video is an OrcaRouter route; the video generation itself stays on the vendor that sells it, and the point is that everything deciding what to generate can live somewhere else.

Choosing, concretely

• Take Seedance 2.5 when the job is a sequence with a cast. Fifty references, standalone audio input, clay-render structural control and a first-place image-to-video position on 6,540 votes describe a model built for continuity work, and $34 to $41 per minute is the price of it.

• Take FLUX 3 Video when the job is resolution or volume. A published 4K rate, a $0.06 Draft tier and a $8.70-per-minute Max tier that ties Seedance on text-to-video are a better fit for high-throughput or high-raster delivery, and you do not need an invitation to buy any of it.

• Take FLUX 3 Max over FLUX 3 if speed is the requirement, but price the claim before you plan around it. Its own board position on image-to-video — sixteenth, 93 points down — is the number to test against your own brief, not the 22×.

• Take neither for editing. FLUX 3 Video has no editing row; Seedance 2.5 does, and Wan 3.0 beats it there at 1,192 against 1,147 while charging $12.00 per minute against $40.82.

What would settle the speed question is a third party running the same prompt at the same resolution and duration on both models and publishing wall-clock times alongside the output. Until someone does, the defensible reading of the October post is that BFL built a genuinely fast tier — a 10-second clip in 12 seconds is plausible for a model documented as handling multi-scene clips and keyframe transitions — and that the multiplier attached to it is marketing arithmetic. Buy FLUX 3 Max because it is fast enough for your loop and cheap enough for your budget. Do not buy it because a footnote-free comparison said it was twenty-two times anything.