Models compared
Every current video model in one table, and which one wins at which job.
Seedance 2.5 is the model creators compare everything else against. Kling 3.0 is the pick for action, Veo 3.1 for polished atmosphere and natural sound, and Wan 3.0 for long clips at a low price. The table below has the details, and the section after it says who wins at what.
Sora 2 is gone. OpenAI shut the Sora app on 26 Apr 2026 and the Sora 2 API on 24 Sep 2026, with no direct replacement. Treat any platform still selling "Sora 2" with suspicion.
Master table
Specs come from official docs. Strengths and weaknesses come from creator tests and community posts, so read them as reported, not measured.
| Model | Max length per clip | Max resolution | Native audio | References | Strengths (reported) | Weaknesses (reported) |
|---|---|---|---|---|---|---|
| Seedance 2.5 | 30 s | 1080p | Yes, dialogue in 11 languages including Vietnamese | Up to 50: 30 images, 10 videos, 10 audio | Taste and complex scenes, wins most side by side tests, long single takes | Strict moderation, no real human faces in references, invents or ignores the audio reference in about 15 to 20% of runs (one tester), some saw a regression from 2.0 |
| Seedance 2.0 | 15 s | 4K (Fast and Mini: 720p) | Yes | Up to 15: 9 images, 3 videos, 3 audio | Only Seedance with 4K, cheaper, beat 2.5 in one same prompt test | Ignores timestamps, only follows "Shot 1, Shot 2" |
| Kling 3.0 | 15 s | 1080p (4K at a higher credit rate) | Yes, mono, 5 dialogue languages | 7 images, or 1 video plus 4 | Action, fights, tracking and orbit shots, multi shot up to 6 shots, Elements for identity | Less predictable between runs, image to video can hang at 99% |
| Kling 4.0 Flash | 20 s | 720p | Not stated | Omni Reference | About 4x cheaper per second than Seedance 2.5 (one tester), dialogue timing praised in posts that look promotional | Only for Ultra yearly subscribers, no first or last frame, generated music sounds dated |
| Veo 3.1 Standard | 8 s (extend to 148 s) | 4K | Yes, always on | 3 images | Polish, physics, ambience and dialogue audio, establishing shots | Pulls stylized characters toward realism, heavy flagging in Flow, expensive |
| Veo 3.1 Fast | 8 s | 4K | Yes | 3 images | Good draft model, 20 Flow credits per clip | Very detailed prompts can produce unrelated output (reported) |
| Veo 3.1 Lite | 8 s | 1080p | Yes | None | Cheapest Google model, 10 Flow credits per clip | No reference images, no 4K |
| Gemini Omni Flash | 10 s (extend to 40 s) | 720p native, 1080p and 4K upscaled | Yes | Reference to video, up to 3 video clips of 3 s | Conversational edits, storyboard grid input, top of an audio leaderboard in Sep 2026 | Native 720p only, not covered by Flow's free daily credits |
| Wan 3.0 | 30 s | 1080p | Yes, free to toggle | Up to 20: 10 images, 5 videos, 5 audio, plus a document or link | Ranked first on Artificial Analysis text to video, fast, cheap | Weak audio, environments can look generated, no longer open weights |
| MiniMax H3 | 15 s | 2K | Yes, stereo | 9 images, 3 videos, 3 audio | Open weights (runs on a 4090 or 5090), H3 Max leads a video editing leaderboard | Blurry mid frames, slow, retries can make it costly |
| Vidu Q4 Preview | 16 s | 4K | Yes, voice cloning from 3 audio clips | 15 images, 3 audio | Many references and 4K on paper | No text only mode, almost all posts are one campaign, no independent review found |
| Runway Gen-4.5 | About 10 s (not confirmed) | 720p generation, 4K upscale (not confirmed) | Not confirmed | First frame | Control and VFX tools, Aleph 2.0 for editing | Little creator talk, Runway lists weak object permanence and cause and effect |
| Luma Ray3.2 | 20 s | 1080p HDR, 16 bit EXR | Not mentioned | Up to 16 keyframes | HDR and EXR for finishing, many keyframes | Almost no community feedback, no sound |
| Grok Imagine 1.5 | 15 s | 1080p | Yes, with lip sync | 14 images, 3 voice references | Fast, Lite tier is very cheap, good for simple loops | Floaty, sluggish movement, weak on complex motion |
| LTX-2.5 | Not stated | 4K | Yes (vendor table) | Not stated | Open weights, free commercial use under 10M USD revenue, useful for upscale, mattes and HDR | Not a hero model for action |
| PixVerse V6 | 15 s | 1080p | Yes, toggle | Fusion, including video references | Cheap, first and last frame, extend | Few independent reviews, vendor says complex action needs retries |
Kling 4.0 (the full model, up to 30 s, 4K, stereo audio, 15 references) was announced for October 2026. On 15 Oct 2026 no official page confirmed it was generally available.
Who wins at what
| Job | Winner | Runner up |
|---|---|---|
| Best overall look, complex scenes | Seedance 2.5 | Seedance 2.0 |
| Action, fights, fast camera | Kling 3.0 | Seedance 2.5 |
| Atmosphere, establishing shots, natural ambience | Veo 3.1 | Gemini Omni Flash |
| Dialogue in Vietnamese | Seedance 2.5 | None confirmed |
| Many reference images (characters, products) | Seedance 2.5 (50) | Wan 3.0 (20), Vidu Q4 (15) |
| One long take | Seedance 2.5, Wan 3.0 (30 s) | Luma Ray3.2 (20 s) |
| Cheap drafts | Veo 3.1 Lite, Seedance 2.0 Mini | Grok Imagine 1.5 Lite |
| Editing a clip you already have | MiniMax H3 Max | Gemini Omni Flash, Runway Aleph 2.0 |
| Running locally, no subscription | MiniMax H3 | LTX-2.5 |
| HDR and finishing | Luma Ray3.2 | LTX-2.5 |
| Free start | Veo 3.1 in Google Flow | Dreamina daily credits |
Pick a model per shot
You do not have to use one model for the whole film. A common split: Kling for action, Veo or Omni for atmosphere, Seedance for shots driven by references or music. Cut them together in the edit.
For prices per clip, see Pricing. For model specific prompting, see Model guides.
Sources
- Official specs: BytePlus ModelArk Seedance docs, kling.ai feature pages and user guide, Gemini API Veo and Omni docs, Alibaba Model Studio Wan 3.0 docs, MiniMax platform docs, Vidu platform docs, Runway research page, Luma Ray3.2 announcement, xAI video docs, ltx.io model page, PixVerse API docs. Links in Sources.
- Strengths and weaknesses: creator tests and posts on X between Aug and Oct 2026, plus practitioner write-ups (Liromfilms, Kapwing, elser.ai). Posts tagged as sponsored were not used, and posts that look like campaigns are labeled.