The video model field rearranged itself this year. OpenAI closed Sora in two stages, the app in April and the API on 24 September, and named no replacement. Every serious model left standing generates sound in the same pass as the picture. Kuaishou opened Kling 4.0 at the end of September and the full model is landing this month. Google retires the three Veo 3.1 preview endpoints on 22 October and points developers at Gemini Omni.
This is the state of things on 15 October 2026, from vendor pricing pages and the model deprecation trackers. Test results on our own briefs follow in later posts.
The roster
| Model | Max clip | Max resolution | Audio | API price | Fits |
|---|---|---|---|---|---|
| Gemini Omni (Omni Flash / Omni 1.1 Flash) | 10s | 720p | Native | Google AI plans; API via Gemini | Everyday social and marketing clips, conversational editing |
| Veo 3.1 (Standard / Fast / Lite) | 4, 6 or 8s | 4K (Standard, Fast) | Native, 48 kHz dialogue | $0.40/s at 720p–1080p, $0.60/s at 4K; Fast $0.10–$0.30/s; Lite $0.05–$0.08/s | Hero shots, dialogue, cinematic deliverables |
| Kling 3.0 / 4.0 | 15s (3.0), 30s (4.0) | 4K, 60 fps on 3.0 | Native, lip-sync in 5 languages | Credits: a 5s 1080p clip with audio runs 60 credits, about $0.80 on Standard | Multi-shot sequences, multi-character dialogue, action |
| Seedance 2.5 | Up to 30s | 4K | Native | Dreamina plans; per-second rates move, check the panel | Long single takes, heavy reference sets (up to 50 inputs) |
| MiniMax H3 | 4–15s | 2K (via regeneration) | Native stereo, 32 kHz | $0.08/s at 768p, $0.13/s at 2K | High resolution per dollar, self-hosted pipelines (open weights) |
| Wan 3.0 | Up to 30s | 1080p | Native | $0.05/s at 480p, $0.10 at 720p, $0.20 at 1080p | Document-to-video, API-first production, itemized budgets |
| Runway Gen-4.5 | About 10s | 4K upscale | Separate audio tools | 12 credits per second, about $0.12/s | Shot control, motion brush, editing after generation |
| Grok Imagine Video 1.5 | 15s | 720p (1080p image-to-video) | Native | $0.08/s at 480p, $0.14 at 720p | High-volume social clips at the lowest rate |
On the Artificial Analysis video arena, the with-audio leaderboard’s top three sit within a few points of each other: Gemini Omni Flash, Alibaba’s Wan 3.0, and fal’s tuned MiniMax H3 Max. Kling 3.0 sits lower on that board and remains the model most teams actually run. Rank and usefulness have stopped tracking each other.
What each one is for
Gemini Omni Flash leads the arena and costs a fraction of the flagships, with two limits: 10 seconds and 720p. Its trick is conversational editing, refining a clip with follow-up prompts instead of regenerating. For social and ad variants, that combination is hard to beat right now.
Veo 3.1 remains the pick when a single shot has to look and sound real on a big screen: native 4K, 48 kHz audio, scene extension past the 8-second base. It is also the one with a clock on it. The preview endpoints shut down on 22 October, so anything you built on them needs a move to Omni this month.
Kling 3.0 is the workhorse. Fifteen second clips, native 4K at up to 60 fps, multi-shot generation with camera cuts inside one pass, and lip-synced dialogue in five languages. Kling 3.0 Omni handles two characters in conversation, each lip-synced to their own lines, and nothing else in the market does that as cleanly. Kling 4.0 doubles the clip ceiling to 30 seconds, adds 10-bit HDR output (listed as coming soon), and takes up to 10 keyframes. Kling has not published 4.0 credit rates yet, and its API still serves 3.0, so budget on 3.0 and treat 4.0 as upside for now.
Seedance 2.5 in Dreamina is the longest single pass you can buy, 30 seconds, with up to 50 image, video and audio references steering the scene. If you are matching a brand kit or extending a product line across shots, that reference depth is the feature. Per-second pricing moves often; the promotional rates that ran through 9 October are gone.
MiniMax H3 sits within a few Elo points of the arena lead at 2K for $0.13 a second, and MiniMax published the weights on 3 August under a community license. If you want to run the model yourself, tune it in ComfyUI, or pin your own rendering stack, this is the strongest option above 1080p.
Wan 3.0 turns a product brief, a document or a web page into a finished audiovisual clip through Alibaba’s API, up to 30 seconds with line-item pricing per resolution. It is the cheapest route to sound-included video at 480p, and the one to pick when the generation has to start from paperwork rather than a prompt.
Runway Gen-4.5 is a studio around a model: camera control, motion brush, character consistency, editing, and ProRes and HDR output on the top tier. It costs more per second than most and earns it when you have to iterate on a shot rather than reroll it.
How to choose
Four rules that survive contact with a client deadline.
- Price the accepted clip, not the generated one. A model at $0.05 a second that needs four retries costs $0.25. Cheapest listed rate is a marketing number.
- Match the ceiling to the format. Ten seconds at 720p is a Reels ad. If the deliverable is a 30-second spot, you are buying Kling 4.0, Seedance 2.5 or Wan 3.0, or you are editing three clips together.
- Check who owns the output. Runway grants commercial use on every plan. Pika and HeyGen reserve it for paid tiers. Free tiers on several platforms are personal use only and stay watermarked. Read the terms for the plan you actually pay for before a clip ships in an ad.
- Do not build on a preview endpoint. Sora’s API died on six months’ notice and Veo 3.1’s previews die on 22 October. Wrap any model in a thin adapter so a swap is an afternoon, not a rewrite.
The default stack
If you shoot marketing and social work and want one subscription plus one cheap API: Gemini Omni Flash through a Google AI plan for drafts and variants, Kling on Standard for anything that needs 4K, multi-shot or dialogue, and Wan 3.0 or MiniMax H3 on the API when volume matters. Add Runway when a shot has to be directed rather than prompted.
Two things to watch before the next update to this page: Kling 4.0 credit rates, and what Google ships next for Omni now that it is the recommended replacement for the retiring Veo previews.
Next in this series: the same brief run through five models, with the outputs side by side. Subscribe below if you want the results.
