Seedance 2 Mini API: Now 50% Off, and the Cheapest Anywhere

Seedance 2 Mini is now 50% off on Segmind, and that puts it at $1.75 per million tokens. The number that matters for most people reading this is the one next to it: fal.ai sells the same model at $7.00 per million. That is exactly 4x our rate, which makes Segmind 75% cheaper for identical output. ByteDance's own platform, BytePlus ModelArk, sits in between at $3.50 per million. Same model, same weights, same token formula on all three.

What Seedance 2 Mini actually is

Seedance 2 Mini is ByteDance's lightweight tier of the Seedance 2.0 family. It is a distilled version of the flagship: the same unified multimodal architecture, running roughly twice as fast as the older Seedance 2.0 Fast tier.

The capability envelope is narrower than flagship Seedance 2.0, and that narrowing is the entire point:

• Resolution: 480p and 720p. No 1080p, no 4k. For those you step up to Seedance 2.0.
• Duration: 4, 5, 6, 8, 10, 12 or 15 seconds. Those are the only legal values, so do not build a UI that lets someone ask for 7.
• Aspect ratios: the full suite. 16:9, 9:16, 1:1, 4:3, 3:4, 21:9 and adaptive.
• Native audio: generate_audio co-generates dialogue, effects and ambience in the same pass. It defaults to true on this model and it costs nothing.
• Multimodal inputs: text alone, a still via first_frame_url, or up to 9 reference images, 3 reference videos and 3 reference audio clips. You cite them in the prompt as image 1, video 1, audio 1.

What a clip costs at every setting

Text-to-video and image-to-video bill at the same rate, so one table covers both. Aspect ratio barely moves the number, because pixel count is what matters and 16:9 and 9:16 are the same pixel count in opposite orientations. My two 720p clips at those two ratios billed identically, to the cent.

A few things fall out of that. A 15 second clip at 720p costs less than 60 cents. Dropping from 720p to 480p cuts your bill by 53%, because 480p 16:9 is a 864 by 496 canvas and that is under half the pixels of 720p. And duration scales linearly once you are past the extra frame, so a 10 second clip is almost exactly twice a 5 second one.

Video-to-video bills more cheaply per token, at $1.05 per million, but the reference video you pass in is counted too. That makes the effective cost per second of output depend on how long and how large your input clip is, so keep reference videos short if you are watching the meter.

Segmind vs fal: the same model at a quarter of the price

Because the token formula is identical everywhere, a cross-provider comparison is pure rate arithmetic. No benchmarking, no quality caveats to argue about. It is the same model producing the same output from the same weights, and the only variable is what each platform charges per million tokens.

fal publishes its rate as $0.007 per 1000 tokens, which is $7.00 per million. We charge $1.75. That is a flat 4x on the headline rate, and because the rate is the only difference, the saving does not move with resolution, duration or aspect ratio. There is no configuration where fal closes the gap.

So the honest per-clip comparison is 74.8% rather than a clean 75%, because that extra frame is the one thing on this model that goes fal's way. I would rather quote their number the way they compute it than inflate it by a percent to make our table look better.

What the saving is worth at your volume

Every $0.5654 you do not spend on a 5 second 720p clip compounds, and short-form video is a volume business. Here is the same clip at four realistic volumes:

Another way to read that table: at what fal charges for one 5 second 720p clip, you get four on Segmind. For a team running a hundred clips a day, the $20,638 a year that does not go to fal would buy nearly three more years of the same output on Segmind.

Going direct to ByteDance does not rescue the comparison either. BytePlus ModelArk is the platform that built the model and it still charges twice what we do, so the cheapest published route to Seedance 2 Mini is not the one closest to the source.

One warning if you are comparing rate cards yourself. fal quotes a per-second summary of $0.1547 per second at 720p, but their own token math works out to $0.1512 per second, so their headline figure overstates their own pricing by about 2.3%. Every fal number in this post uses the cheaper, formula-derived figure. Price from the token formula, not from a vendor's rounded per-second number, including ours.

Use case 1: vertical product ads for a marketing agency

The workload that pays for itself fastest here is ad variant generation. An agency running paid social needs the same product in eight framings with four camera moves, and it needs them this afternoon. At 19 cents a clip you stop rationing attempts and start actually testing.

I ran a clean studio product shot at 720p in 9:16, the Reels and TikTok frame.

The product holds its shape through the orbit, the key light stays put, and the shadow tracks correctly under the pedestal. That is the bar that matters for product video: not cinematic flair, but the object not deforming while the camera moves. The 720x1280 container came back correctly oriented too, which is worth checking on any vertical generation.

Second, iterating here is genuinely trivial. Eight variants of this shot at 720p and 5 seconds is $1.52.

The API call

import requests

r = requests.post(
    "https://api.segmind.com/v1/seedance-2.0-mini",
    headers={"x-api-key": "YOUR_API_KEY"},
    json={
        "prompt": "A pair of white running shoes rotating slowly on a matte grey pedestal...",
        "duration": 5,
        "resolution": "720p",
        "aspect_ratio": "9:16",
        "generate_audio": True,
    },
)
open("ad-variant.mp4", "wb").write(r.content)
print("billed:", r.headers["x-cost"])

That last line is the habit I would most like to pass on. Log x-cost on every call. It is the only real source of truth for what you spent, and it is how every number in this post was checked.

Use case 2: previz and shot exploration for a film studio

Previsualisation is the other place these economics change behaviour. A director wants to see whether a shot reads before anyone books a location. At 19 cents you can explore fifteen versions of a scene in an afternoon for under three dollars, which is less than the cost of the meeting where you would otherwise have argued about it.

For a 19 cent draft this is a genuinely usable previz plate. The wet cobblestones carry the reflected signage, the rain reads at the right scale against the figure, and the single overhead lamp does real volumetric work. The handheld sway I asked for is present and restrained rather than seasick. At 720p you would not cut this into a finished film, but that is not what previz is for: it answers "does this blocking work" for the price of nothing.

Two practical notes on prompting this tier. Describe one continuous scene with one main action rather than a shot list, unless you are at 12 seconds or longer where each shot has room to breathe. And keep audio atmospheric rather than per-object. Asking for specific foley on individual props is a reliable way to get a 500 back.

Use case 3: volume drafting for production houses and MCNs

If you are producing hundreds of clips a month, 480p is where you live. It is a 864 by 496 canvas at 10,044 tokens per second, under half the tokens of 720p, for output that is perfectly adequate for internal review, animatics, and deciding which concepts deserve a real render.

I fired the cheapest legal configuration this model offers, a 4 second 480p clip at 7 cents.

Then the long end, 15 seconds and vertical, which is the shape of a full short-form beat rather than a single shot. Fifteen seconds of finished video with synchronised audio for 26 cents.

The workflow I would actually recommend for volume is the one this model is designed around. Draft everything at 480p, review, then re-render only the approved shots, taking the ones that need real fidelity up to flagship Seedance 2.0 at 1080p. You pay full price only for the shots that survived review, which on a typical campaign is a small fraction of what you generated.

Controlling spend in production

Three parameters do most of the work on your bill, and one does none of it.

resolution is the biggest lever. 480p to 720p is a 2.15x jump in tokens. Default your draft path to 480p and make 720p an explicit choice.

duration is linear, so there is nothing clever to do except not ask for 15 seconds when 5 answers the question.

aspect_ratio barely matters. 1:1 at 480p is a 640 by 640 canvas and marginally the cheapest shape available, but the spread across ratios at one resolution is a few percent. Choose by platform, not by price.

bitrate_mode does not affect price at all. Setting it to high gives you roughly five to six times the bitrate for the same money, which is free quality if you are going to grade or re-encode. It is the one setting with no trade-off to think about.

Two more habits worth building in. Log the x-cost header on every call, as above. And note that failed generations bill nothing, so iterating against parameter validation errors costs you latency and not money.

Honest assessment

What this model is very good at: cost per attempt, and predictability. Native audio in the same pass, for free, removes an entire step from a short-form pipeline. And I can quote any configuration to the cent without firing it, which is rarer than it should be.

Where it falls short: 720p is the ceiling, so it is not a delivery model for anything that needs to hold up on a large screen. Real human faces are blocked, which rules out a whole category of UGC and testimonial work. And the wall clock is variable: my five clips ranged from 90 to 149 seconds with no clean relationship to duration, so I would not promise a completion time inside a user-facing flow. Treat latency as a distribution, not a number, and make the call asynchronous.

Best fit: high-volume short-form, product and e-commerce video, ad variant testing, previz and animatics. Not a fit: final delivery above 720p, anything needing a real person's likeness, or a synchronous UI that blocks on the render.

FAQ

How much does the Seedance 2 Mini API cost?

$1.75 per million tokens for text or image input during the current limited time offer, and $1.05 per million for video input. In practice that is $0.1906 for a 5 second 720p clip and $0.0886 for the same clip at 480p. Audio is free.

Is Seedance 2 Mini cheaper on Segmind than on fal.ai?

Yes, by 75% on the published rate: $1.75 per million tokens on Segmind against $7.00 on fal.ai, for the same model and the same token formula. A 5 second 720p clip is $0.1906 here and $0.7560 there. At 100 clips a day that is $1,696 a month, or $20,638 a year, that you keep. The saving is the same percentage at every resolution, duration and aspect ratio, because the rate is the only thing that differs.

What is the cheapest way to use Seedance 2 Mini?

Segmind at $1.75 per million tokens is the lowest published rate I could find for this model anywhere, a quarter of fal's and half of ByteDance's own. Within Segmind, 480p is 53% cheaper than 720p for the same duration, and generating at the shortest duration that answers your question is the other main lever.

Is Seedance 2 Mini cheaper than Seedance 2.0?

Yes. It bills the same token count for a given clip but at a lower per-token rate. It also caps at 720p where flagship Seedance 2.0 reaches 1080p and 4k, so the saving comes with a real ceiling on resolution.

Does Seedance 2 Mini generate audio?

Yes, natively and in the same pass as the video. generate_audio defaults to true on this model and adds nothing to your bill. Describe ambience rather than asking for music, which can trip an output-side content check.

How long can a Seedance 2 Mini clip be?

4 to 15 seconds, but only at the discrete values 4, 5, 6, 8, 10, 12 and 15. Any other number is rejected, so validate before you call.

Can Seedance 2 Mini generate real people?

No. Face blocking applies across every Seedance tier, so a first frame or reference image containing a recognisable real human face is rejected rather than degraded.

The short version

Seedance 2 Mini is the tier you reach for when the number of attempts matters more than the ceiling on any one of them, and at $1.75 per million tokens the cost per attempt is now low enough that rationing them is the wrong instinct. A 5 second 720p clip with synchronised audio is 19 cents, verified against the billed header rather than a rate card. The same clip is 38 cents on ByteDance's own platform and 76 cents on fal.ai, so the entire decision is whether you would rather pay 4x for identical frames.

If you are already generating on fal, the switch is worth $1,696 a month at a hundred clips a day and $20,638 over a year, and nothing about your output changes. The offer is time limited. If you are running short-form video at any real volume, this is the window. Try Seedance 2 Mini on Segmind, and log the x-cost header on your first call so you can watch the arithmetic work for yourself.