The benchmark says ByteDance won
On July 31, 2026, ByteDance quietly dropped Seedance 2.5, and if you only read the spec sheet, it's not close. Native 30-second clips in a single pass (competitors top out around 10-15 without stitching). Up to 50 multimodal reference inputs — 30 images, 10 videos, 10 audio clips — fed into one generation. Timestamp-level editing so you can swap a background at second 12 without re-rendering the whole clip.
Its predecessor, Seedance 2.0, already sits at the top of the Artificial Analysis Video Arena: Elo 1,269 for text-to-video, 1,351 for image-to-video — first place in both categories, ahead of Kling 3.0 and Google Veo 3. Seedance 2.5 is the model built on top of that lead.
On raw capability, here's how the field actually stacks up right now:
| Model | Strength | Weak spot |
|---|---|---|
| Seedance 2.5 | Duration, reference count, brand/character consistency | No public API at launch |
| Sora 2 | Physics simulation — fluids, gravity, structural deformation | Shorter clips, tighter compute limits |
| Veo 3.1 | Native audio generation, actual developer access | Loses on reference flexibility |
| Kling 3.0 | Resolution, has a free tier | Loses the Elo race outright |
If you're picking a model purely on "what can it generate," Seedance 2.5 is the correct answer today. That's not really in dispute.
Here's the part nobody's pricing sheet mentions
You can't build anything with it. Not yet.
Seedance 2.5 rolled out on Jimeng AI and Doubao Pro — ByteDance's own consumer apps. If you're a developer who wants to wire this into a product, your options are: wait, or route through a third-party reseller scraping the consumer UI. ByteDance says API access is coming "soon" via BytePlus ModelArk and Volcano Engine. No date. We've heard "soon" before — Seedance 2.0's API took months to catch up to its consumer release.
This is the same playbook every frontier lab runs now: ship the demo that wins the leaderboard and the Twitter cycle, gate the thing developers actually need behind a waitlist. Sora did it. Veo did it less aggressively because Google needed the Vertex AI story. ByteDance is doing it with the added twist that the flagship consumer surfaces are Chinese apps most Western dev teams have never opened.
What the eventual API will probably look like
Volcano Engine's pricing already leaked ahead of the API: ¥42 per million billable tokens with video input, ¥70 without (roughly $6.40 / $10.70 on BytePlus's dollar pricing). Rough math puts a 30-second clip at $0.66–$1.80 on standard settings, climbing past $15 at 4K. That's steep for anything you'd run at scale — think batch content generation, not a per-request user feature.
Given how ByteDance's other models expose generation (Seedance 2.0, Doubao), expect an async job pattern, not a synchronous call. Something like:
import requests, time
resp = requests.post(
"https://ark.byteplus.com/api/v3/video/generations",
headers={"Authorization": f"Bearer {API_KEY}"},
json={
"model": "seedance-2-5",
"prompt": "a drone shot pulling back from a rooftop garden at dawn",
"duration": 30,
"references": {"images": [...], "videos": [...]},
},
)
job_id = resp.json()["id"]
while True:
status = requests.get(f".../video/generations/{job_id}",
headers={"Authorization": f"Bearer {API_KEY}"}).json()
if status["state"] in ("succeeded", "failed"):
break
time.sleep(5)
Long-running video jobs almost never come back synchronously at this duration and reference count — the compute cost per request is too high to hold a connection open. If you've integrated Sora or Runway, this shape will feel familiar: submit, poll, pull the asset URL, pay per token or per second.
The actual takeaway for builders
Don't pick a video model off a leaderboard. Pick it off what you can ship this quarter.
If you need something in production now, Veo 3.1 is the boring, correct choice — real API, real audio, real docs. If you're prototyping something that lives entirely inside the demo-and-fundraise phase and can tolerate waiting on ByteDance's access rollout, Seedance 2.5's ceiling is genuinely higher and worth the wait. Everyone else chasing "best benchmark" is optimizing for a metric that doesn't compile into a working product.
The Elo score is real. The moat is real. The API is not — yet. Build accordingly.
Top comments (0)