Gemini Omni 1.2 Flash vs Seedance 2.5: Price per Second, Limits and Which to Use

Compare Gemini Omni 1.2 Flash and Seedance 2.5 on clip length, resolution, aspect ratio, audio, inputs, API flow and official per-second prices, with a test plan.

Two paper kites on strings of different lengths, one flying close and one flying far, both tied to a single orange reel resting on the grass.

Gemini Omni 1.2 Flash is the cheaper choice for short 16:9 or 9:16 clips up to 15 seconds: at Google’s list price a 720p second costs about $0.07, against about $0.23 for Seedance 2.5 at ByteDance’s list price. Seedance 2.5 is the one to pick when a single request must run longer (up to 30 seconds), use a square, 4:3 or 21:9 frame, accept audio references, or carry dialogue in a language Google has not evaluated for Omni (Google fully supports only English). Neither model wins on every brief. The deciding factors are clip length, aspect ratio, the inputs you need to send, and where you are allowed to run the model.

This comparison is built from vendor documentation rechecked on November 19, 2026: Google’s Omni guide, model card and pricing page, and ByteDance’s BytePlus ModelArk Seedance 2.5 tutorial and pricing page. The per-second numbers are our arithmetic on those published rates, shown step by step.

Ofox publishes this article, and Seedance 2.5 is available through Ofox while Gemini Omni is not. We ran two real Seedance 2.5 generations through Ofox on November 19, 2026. We did not run Omni, because we had no Google project for it, so this article makes no quality claim about which model produces better video. What it gives you instead is a same-basis spec and price comparison, and a test plan for running both models on your own briefs.

Spec sheet: what each model can produce

Both models generate video with sound from text, images or video. The differences are in the limits.

DimensionGemini Omni 1.2 FlashSeedance 2.5
Model IDgemini-omni-1.2-flashdreamina-seedance-2-5-260628 (ModelArk), bytedance/seedance-2.5 (Ofox)
New clip length3–15 s, whole seconds; 10 s default4–30 s, whole seconds, or -1 to let the model choose; -1 default
First + last frame clip3–10 s4–30 s range; ratio must be adaptive
Extension+3–10 s per turn, up to 30 s total, appends to the end onlySupported (reference video plus extension intent); ratio must be adaptive
Resolution360p, 720p (default), 1080p, 4k; Google labels 1080p and 4K as upscaled480p, 720p (8-bit), 1080p, 4k (10-bit, H.265/HEVC); 720p default
Frame rate24 fpsByteDance’s pricing examples reconcile at 24 fps; our outputs were 24 fps
Aspect ratio16:9 (default) or 9:1621:9, 16:9, 4:3, 1:1, 3:4, 9:16, adaptive (default)
Audio outputNative audio, generated by defaultSynchronized audio, generated by default
Audio inputNot supported: no audio inputs or audio referencesUp to 10 audio references, 30 s total
Image referencesUp to 6 (up to 3 per character, up to 2 characters)Up to 30
Video referencesUp to 3 (clips of 3–10 s) or 1 (clips of 11–15 s), each up to 3 s; ignored audioUp to 10, 2–30 s each, 30 s total
EditingConversational, over turns with previous_interaction_id; input video ≤15 sReference video plus edit wording; source 4–30 s; output keeps its length and ratio
Dialogue languagesEnglish fully supported; others “have not been evaluated”11 languages listed, including English, Chinese, Spanish, Japanese and Korean
Output containerMP4MP4 (default) or MOV
Draft previewNot documentedDraft mode: fixed 480p preview you can promote to a final render

Two rows decide most choices before price comes into it. Omni only makes landscape or portrait video, so a square feed post or a 21:9 banner rules it out. And Omni’s longest single generation is 15 seconds. You can reach 30 seconds through extension turns, but each turn is a separate request that looks back at only the last 10 seconds of the clip, which is a different workflow from writing one 30-second timeline.

The resolution row needs a careful reading. Google says its 1080p and 4K outputs are produced by upscaling. ByteDance documents 1080p and 4K for Seedance 2.5 as 10-bit HEVC output and does not describe them as upscaled. The vendors do not use the same words for the same thing, so the 1080p and 4K rows in the price table below are not like-for-like products. Only judge them on frames you have looked at yourself.

If you have read our Seedance 2.5 prompting guide (published August 14, 2026), note that the spec has moved on since then. That guide reported a 480p/720p ceiling, because ByteDance’s tutorial at the time (updated August 13) still listed 1080p and 4K as unsupported and 4K as a Seedance 2.0 feature. The tutorial as of its November 16, 2026 update lists 1080p and 4K for Seedance 2.5, and a cheaper sibling, Seedance 2.5 Flash (dreamina-seedance-2-5-flash-261115), which is limited to 480p/720p and 20 reference assets. The Ofox catalog also lists bytedance/seedance-2.5-flash as of November 19, 2026; we did not test it.

What a second of video costs at the same resolution

Both vendors bill in tokens, but their tokens are not the same unit. You cannot compare “$12 per million” with “$10.70 per million” directly. The only fair basis is dollars per second of output, at the same resolution, aspect ratio and input mode.

Google gives the conversion itself. Omni 1.2 Flash video output is $12.00 per 1M tokens on the paid tier (there is no free tier), and the pricing footnote fixes the tokens per second: 1,931 at 360p, 5,792 at 720p, 8,688 at 1080p and 17,376 at 4K. Per-second price = tokens per second × $12 / 1,000,000.

ByteDance publishes a formula: tokens = (input video duration + output duration) × width × height × frame rate / 1024. With no video input, the Seedance 2.5 rate is $10.70 per 1M at 480p/720p, $11.70 at 1080p and $4.70 at 4K. To turn that into a per-second price we assume no video input, 24 fps, and the pixel sizes ByteDance’s tutorial lists for 16:9: 854 × 480, 1280 × 720, 1920 × 1080 and 3840 × 2160. A 720p 16:9 second is then 1280 × 720 × 24 / 1024 = 21,600 tokens, or 21,600 × $10.70 / 1,000,000 = $0.2311. A 9:16 frame swaps width and height, so it uses the same token count; the other listed ratios land within about 1% of the 16:9 pixel count. ByteDance calls these estimates and bills the usage.completion_tokens its API returns.

ResolutionOmni 1.2 Flash per secondSeedance 2.5 per secondSeedance ÷ Omni
360p (Omni only)$0.0232not offeredn/a
480p (Seedance only)not offered$0.1028n/a
720p$0.0695$0.23113.3×
1080p$0.1043 (upscaled)$0.56865.5×
4K$0.2085 (upscaled)$0.91374.4×

These are list prices calculated from the vendors’ own pages on November 19, 2026 (ByteDance’s pricing page was last updated November 16), with text or image input and no video input. Our arithmetic matches both vendors’ published examples. Google quotes ”~$0.70 for a 10s clip, ~$1.04 for a 15s clip” at 720p, and about $0.023, $0.104 and $0.209 per second at 360p, 1080p and 4K. ByteDance prices a 5-second 16:9 clip at $0.514 (480p), $1.156 (720p), $2.843 (1080p) and $4.568 (4K), which is our per-second figure times five.

The 720p row is the cleanest comparison, because both vendors make 720p natively. Here Omni costs about a third of Seedance per second.

Clip prices that match real briefs:

ClipOmni 1.2 FlashSeedance 2.5
5 s, 720p$0.35$1.16
10 s, 720p$0.70$2.31
15 s, 720p$1.04$3.47
30 s, 720pThree requests (15 + 10 + 5 s); total not computable from the published page$6.93 in one request
5 s, 480p draft-quality360p: $0.12480p: $0.51

The 30-second Omni row is blank on purpose. Each extension turn produces new output tokens, and Google lists video input at $1.50 per 1M tokens. But the pricing page does not say how many input tokens a second of context video counts as, or exactly how an extension turn is metered. Until you have measured the usage your own extension turns report, treat that cost as unknown. Do not assume it is the sum of three short-clip prices.

Four terms the per-second table leaves out:

  1. Omni’s other token lines. Prompt text and reference images are input at $1.50 per 1M tokens. Output text, including thinking tokens, is billed at $9.00 per 1M. The REST response can contain a thought step. These amounts are probably small next to the video tokens, but read them from the usage on your first real run instead of assuming zero.
  2. Seedance’s video-input rate. When the input includes a video, the rate drops to $6.40 per 1M (480p/720p), but the reference video’s duration is added to the token count, and ByteDance applies a minimum token charge. Its own example prices a 5-second 720p output at $1.244 to $4.838, depending on how long the input video is.
  3. Failures. ByteDance states that you are charged only for successfully generated videos, including when moderation blocks the request. We found no equivalent statement on Google’s Omni pricing page.
  4. Getting started. On ModelArk, Seedance 2.5 can be activated once any one of these holds: a Pay by Credits balance of at least USD 30, a dedicated AI Savings Plan of USD 30 or more with balance left, or a Seedance 2.0/2.5 resource pack of USD 30 or more with quota left. Our August guide described only the resource-pack route; ByteDance’s resource-pack page, checked November 19, 2026, lists all three. A Seedance 2.5 pack is $32 for 5 million tokens and expires 90 days after purchase. Omni requires a Google project on the paid tier.

Through a gateway, the per-second price is whatever that gateway sets. Ofox shows its current Seedance 2.5 rate on the model page, and billing rules are in the pricing docs. This article does not quote Ofox prices. If you want to audit any reseller against ByteDance’s formula, the method is in our Seedance 2.5 price audit.

Price per second is also not price per usable clip. If a cheaper model needs four attempts to produce a keeper and a dearer one needs one, the ranking flips. Our usable-clip cost guide covers how to count those attempts.

Inputs, references and editing work differently

Omni works as a conversation. You generate a clip with store=True, then edit it in later turns by sending previous_interaction_id and an instruction such as “make the daylight warmer, keep everything else the same”. Google’s prompt guide recommends that phrase for edits. The model infers the task from your prompt. A task field (text_to_video, image_to_video, reference_to_video, edit, extend) exists, but Google calls it a fallback because it adds stricter constraints. Old control tags such as <FIRST_FRAME> are deprecated.

Seedance infers the task from the assets you attach and the words you use. A video tagged reference_video becomes an edit only if the prompt shows editing intent. Otherwise it becomes a reference-driven new clip. Edit, extension and first/last-frame tasks lock ratio to adaptive, and edits also lock duration to -1. If you set omni_reference_task_type explicitly, the constraint check happens when the task is submitted, not after it starts. Our first and last frame guide shows the 400 error you get when you break this rule through Ofox.

Practical consequences:

  • Product consistency across shots. Omni accepts up to 6 reference images. Seedance accepts up to 30, but ByteDance’s own prompt guide says 1–8 image subjects work best and stability may drop at 9–12. For one product and one hero image, either limit is plenty.
  • Driving the voice or the music. Only Seedance accepts audio. Omni cannot take an uploaded voice track or music bed, and voice editing is not supported.
  • Editing footage you shot yourself. Both support it. Omni does not allow editing or extending uploaded videos for users in the EEA, Switzerland or the UK (generated videos are fine). Seedance accepts source videos of 4–30 s, up to 200 MB, at 24–60 fps.
  • Through a gateway. The reference budget is whatever the gateway passes through. Ofox’s video-create reference documents up to 9 images, 3 audio clips (15 seconds each at most) and 1 video in input_references as of its November 18, 2026 update, which is below ModelArk’s ceilings. Going over returns 400 too_many_references. Check this before you design a 20-reference pipeline.

API shape: one call versus a task you poll

The integration patterns differ more than the parameter names do.

Omni runs through Google’s Interactions API. Google recommends background=false, store=false and stream=false for the fastest one-shot generation, which is a synchronous call. Small videos come back inline as base64. For anything over 4 MB, Google recommends delivery="uri": you get a file URI back and poll that file until it is ACTIVE, then download it. Choosing store=false means the clip cannot be edited in a later turn. Our Omni API guide has a complete runner with recovery handling. The shape of the call, on Google’s native API:

from google import genai

client = genai.Client()  # GEMINI_API_KEY from your Google project; Omni is not on Ofox
interaction = client.interactions.create(
    model="gemini-omni-1.2-flash",
    input="A matte white ceramic mug on a pale oak table, soft window light...",
    response_format={"type": "video", "duration": "5s",
                     "resolution": "720p", "aspect_ratio": "9:16", "delivery": "uri"},
    store=True,
)

Seedance is an asynchronous task. On ModelArk you create a generation task and poll the task by ID. Through Ofox it is POST /v1/videos, which returns 202 with an id and a polling_url. You then call GET /v1/videos/{id} until the status is terminal, or you pass a callback_url. The retrieve reference lists the states and the usage object. This is the exact body we sent for one of our test runs:

import os, time, requests

BASE = "https://api.ofox.ai/v1"
H = {"Authorization": f"Bearer {os.environ['OFOX_API_KEY']}"}
body = {
    "model": "bytedance/seedance-2.5",
    "prompt": ("A matte white ceramic coffee mug on a pale oak table, vertical product shot, "
               "soft morning window light from the left. 0-2s: steam rises slowly from the mug, "
               "camera static. 2-5s: the camera pushes in slowly toward the mug handle. "
               "The mug shape, color and position stay the same throughout. "
               "Sound: quiet room tone and one soft ceramic tap at 2s. "
               "No dialogue, no music. No subtitles."),
    "duration": 5, "resolution": "480p", "aspect_ratio": "9:16",
    "generate_audio": True, "seed": 20261119,
}
job = requests.post(f"{BASE}/videos", json=body, headers=H).json()
while True:
    s = requests.get(f"{BASE}/videos/{job['id']}", headers=H).json()
    if s["status"] in ("completed", "failed", "cancelled", "expired"):
        break
    time.sleep(10)
print(s["status"], s.get("usage"), (s.get("mirror_urls") or s.get("unsigned_urls")))

Retention matters for both. ByteDance keeps a Seedance video URL for 24 hours (and a maximum of 100 downloads). A task ID is valid for 7 days, and a Draft task ID for 14 days. Ofox’s unsigned_urls can also expire within 24 hours, so download the file promptly. If your polling loop hits timeouts or long waits, our video polling guide and Ofox’s video error codes cover the failure states. Google-side authentication and quota errors follow the same patterns as in our Gemini 4 API error guide.

Access, rate limits and the December 3 routing change

Gemini Omni 1.2 FlashSeedance 2.5
First-party APIGemini API, paid tier onlyBytePlus ModelArk (activation threshold USD 30)
On Ofox (checked 2026-11-19, GET /v1/models)NoYes: bytedance/seedance-2.5, 4–30 s (see note below on resolution)
Published rate limitsTier 1: 8 RPM / 40 RPD; Tier 2: 16 / 200; Tier 3: 40 / 2,000; Free: not supportedModelArk individual: 180 RPM, 3 concurrent tasks; enterprise: 600 RPM, 10 concurrent; 4K: 15 RPM, 1 concurrent

Omni cannot be called through Ofox today, so every Omni example in this article uses Google’s native endpoint.

One catalog detail, as of November 19, 2026: in Ofox’s GET /v1/models entry for bytedance/seedance-2.5, the video_attributes.resolutions field lists 480p, 720p and 1080p, while the same entry’s text description says 480p/720p. 4K is not listed in either. We generated only at 480p, so we have not checked which resolutions the gateway accepts today; confirm on the model page or with a small request before planning around 1080p. For Seedance through Ofox, the limits are Ofox’s, not ModelArk’s. The current default is in the rate-limit guide.

The Omni daily cap matters for batch work. Forty requests per day on Tier 1 is enough for prototyping, but not for producing a few hundred social variants. Seedance’s concurrency cap acts more like a queue: tasks above the limit wait rather than fail.

On December 3, 2026, Google will route every request for gemini-omni-1.1-flash to gemini-omni-1.2-flash and bill it at 1.2 rates. Video output drops from $17.50 to $12.00 per 1M tokens, and the maximum new-clip length rises from 10 to 15 seconds. If you are benchmarking Omni against Seedance with old 1.1 numbers, re-run them on 1.2: the old prices overstate Omni’s cost, and the old 10-second cap understates its length. Pin gemini-omni-1.2-flash explicitly rather than relying on the alias.

Content rules, people and provenance

Provenance marking. Google states that every Omni video carries an invisible SynthID watermark that can be detected programmatically. You cannot turn it off. Seedance has a watermark parameter on ModelArk that adds a visible AI-generated mark in the lower-right corner. It defaults to false. The Seedance 2.5 tutorial does not document an invisible watermark either way. Ofox’s video-create API has no top-level watermark field, and our two Ofox outputs had no visible mark. If your platform requires AI-content labelling, apply it in your own pipeline rather than relying on either vendor.

Real people. This is the biggest constraint for talking-head work.

  • Seedance 2.5 does not accept uploaded reference images or videos containing real human faces. ByteDance’s routes around this are a library of preset digital characters, authorised real-person assets, and reuse of face-containing outputs generated under the same account.
  • Omni does not support uploading or editing images of “certain recognizable people”. In the EEA, Switzerland and the UK it also refuses images containing minors.
  • Both apply safety filters to prompts and outputs. Google notes that its filters vary by region.
  • Through Ofox, the create reference (updated November 18, 2026) documents a top-level real_person flag that applies privacy-preserving preprocessing to reference images of identifiable people before the task is submitted. We did not test it.

In practice, a talking head of an invented presenter from a text prompt works on both. A talking head of your actual founder from a photo is blocked on Seedance without the authorised-asset process, and may be blocked on Omni.

What our two Seedance runs show

We generated two clips through bytedance/seedance-2.5 on Ofox on November 19, 2026, 03:35 UTC. Settings: 9:16, 480p, 5 seconds, generate_audio: true, seed: 20261119. These were single runs (n = 1 per brief). They show the workflow and billing behaviour, not a quality rate.

Run A: product shotRun B: talking head
BriefWhite mug, steam 0–2 s, slow push-in 2–5 s, ceramic tap, no dialogueInvented presenter, static medium close-up, one English line
Submit → completed120.9 s123.6 s
Billed (usage)5 video seconds5 video seconds
File (ffprobe)480×854, H.264, 24 fps, 5.04 s video; AAC 32 kHzSame
What we checkedFrames at 0.5 s, 2.5 s and 4.5 s: same mug shape and colour, steam present, visible push-in between framesTranscribed with openai/gpt-4o-mini-transcribe: exactly “Here is the one setting most people forget to change.” Mouth moves across sampled frames
Not checkedWhether the ceramic tap lands at 2 sFrame-accurate lip sync

Both clips did what the brief asked on the points we could measure. The audio track in the product shot was quiet (mean −40 dB), which fits “quiet room tone”. The talking head was clearly audible (mean −20 dB).

This tells you Seedance 2.5 handled two common short-video briefs through Ofox in about two minutes each. It tells you nothing about how Omni would do on the same briefs, which is why the next section exists.

Choose by use case

Use caseLean towardsWhyWatch out for
Product shots, 5–10 s, 16:9 or 9:16Omni, if you can use Google’s APIAbout a third of Seedance’s price per second at 720p; 6 image references cover one productNo 1:1 or 4:5; check the product’s text and logos on the actual frames
Product shots for square feeds or 21:9 bannersSeedance 2.5Omni has no 1:1, 4:3 or 21:9Price per second
Talking head, invented presenter, EnglishEither, so test bothBoth generate dialogue with native audioOmni rejects uploads of some recognisable people; neither model’s lip-sync is verified here
Talking head, non-English, or driven by your own voice trackSeedance 2.511 documented languages; audio referencesReal-face uploads need ByteDance’s authorised-asset route
Social vertical, high volumeOmni on price; Seedance on throughputOmni 720p is about $0.07/s, but Tier 1 allows 40 requests a dayCheck your Google tier before planning a batch
Multi-shot story, 15–30 sSeedance 2.5One request with a 30-second timeline; no 10 s lookback$6.93 for a 30 s 720p clip at list, so draft at 480p or in Draft mode first
Multi-shot story, 15 s or lessEitherOmni supports timecode prompts such as [0-5s] ... [5-10s] ...; Seedance supports integer-second timestampsRe-check continuity at each cut
Iterative edits on one generated clipOmniConversational edits over stored interactionsstore=True is required; uploaded-video edits are blocked in the EEA, Switzerland and the UK
You need to call it through OfoxSeedance 2.5Omni is not on Ofox as of 2026-11-19Gateway reference limits are lower than ModelArk’s

For a broader map of which video model fits which job, see how to choose a video generation API by use case. If you are also weighing Alibaba’s model, read Wan 3.0 vs Seedance 2.5.

Run your own head-to-head

A fair test holds everything constant except the model. Here is the plan we would run with access to both, and the one you can run.

Fixed settings. 720p, 9:16, 5 seconds, audio on, no reference video. Omni: resolution: "720p", aspect_ratio: "9:16", duration: "5s". Seedance: resolution: "720p", aspect_ratio: "9:16", duration: 5. Both make 720p natively, and 9:16 is the only shared portrait ratio. At list prices, one clip per brief costs about $0.35 on Omni and $1.16 on Seedance.

Four briefs, one per use case. Use identical wording on both. Write timing in a form both vendors document (0-2s: style), and constraints as plain sentences, because Omni has no negative-prompt field.

  1. Product: the mug prompt from the code block above, at 720p.
  2. Talking head: the invented-presenter prompt with one English line, no subtitles.
  3. Vertical social: “A runner ties a shoe on a city step at sunrise, then stands and starts jogging out of frame. 0-2s close-up on hands, 2-5s wide shot. Street ambience, no music, no text.”
  4. Multi-shot: three timed beats in 5 seconds with a hard cut between each, the same character in a red jacket in every beat.

Score each clip against the brief, not against the other model.

CheckHow to measurePass
File specffprobe: dimensions, fps, video and audio durationsRequested size; video track within 0.25 s of 5 s; audio track present
Subject heldFrames at 0.5 s / 2.5 s / 4.5 sSame object or person, no shape or colour drift
TimingFrames at each beat boundaryEach beat visible in its window
DialogueSpeech-to-text on the extracted audioExact line; no extra speech in non-dialogue briefs
Unwanted elementsWatch the full clipNo captions, music or extra subjects that the brief excluded
Cost and timeVendor usage field; submit-to-file wall clockRecorded, not estimated

Run at least three seeds or attempts per brief before calling a winner. One clip per model is an anecdote, as our own two runs are. Count failed and rejected attempts in the cost column, because the keeper rate is what decides the real price.

Sources