The Newest AI Video Models (2026): Omni Flash, Wan 3.0, Seedance 2.5, Hailuo 2.3 Compared

Four video model families shipped or moved fast in the last few weeks, and all of them are live in MarketDragon's model picker today. This is a working comparison, not a ranking — what each one is reported to be good at, its real limits from our own config, and which one to reach for depending on the shot.

None of this is a rendering benchmark. A prompt here gives you a storyboard shot, not a finished film — you still pick what to render. Where a claim comes from a vendor's own page it says so; where it's a third party's report it says "reported."

For a Filipino creator posting to Facebook Reels or building a Shopee/Lazada product ad, the model you pick changes what you can promise a client: a longer single clip, a steadier face across cuts, or a faster turnaround when a batch of five reels is due today, not next week.

Gemini Omni Flash 1.1 (Google)

Google's Gemini Omni 1.1 Flash became generally available August 27, 2026, as a production-ready update to the Gemini API's video model, with scene extension, first/last-frame control, and video-reference inputs for keeping a character or a motion style consistent across a continuation (Google Developers Blog, source below).

In MarketDragon, google/gemini-omni-flash-1-1 (text-to-video) and its image-to-video sibling are both active. Per config/kie-models.php: prompt cap 20,000 characters, duration 4/6/8/10 seconds, resolution 360p/720p/1080p/4K, aspect 16:9 or 9:16, and a reference-video slot (one clip, up to 30 seconds, a 10-second context window) for carrying a look into a new shot. Audio is supported.

Google's own write-up describes scene extension analyzing up to 10 seconds of prior context to stitch continuations up to 40 seconds total — useful for a beat that runs long, since MarketDragon still renders and stores one clip per beat.

The Wan family (Alibaba)

Wan moved through several versions this year. In MarketDragon, four generations are active at once, each with a different shape:

• Wan 3.0 / Wan 3.0 Prime — the newest. Alibaba launched Wan 3.0 on August 24, 2026 and Wan 3.0 Prime three days later as a faster tier (reported by Dataconomy and OpenRouter's model listing, sources below). In our config, both take a 20,000-character prompt, resolutions 480P/720P/1080P, an adaptive aspect option plus five fixed ratios, and any duration from 2 to 30 seconds. Image-to-video takes either a first/last frame pair or up to 10 reference images — not both in the same render. Audio is supported on both tiers.
• Wan 2.7 — reported to add first/last-frame control and multi-reference video inputs over 2.6 (Oakgen.ai, Kie.ai market page). Our config caps the prompt at 5,000 characters, duration on a 2-15 second ladder, resolution 720p/1080p, and a prompt-rewrite toggle. Image-to-video takes a start frame and an optional end frame.
• Wan 2.6 — reported for affordable multi-shot 1080p with native audio (Kie.ai). Our config: 5,000-character prompt, duration 5/10/15 seconds, 720p/1080p. Its image-to-video variant takes exactly one input image, not a first/last pair.
• Wan 2.5 — the shortest prompt cap in the family at 800 characters in our config, duration 5 or 10 seconds, 720p/1080p, aspect 16:9/9:16/1:1.
• Wan 2.2 Text-to-Video / Image-to-Video Turbo — the budget tier: 480p or 720p only, a fixed ~5-second clip (81 frames at 16fps, per our pricing comment), no duration knob.
• Wan 2.2 Speech-to-Video — a different product surface: a photo plus an audio clip, animated to speak. It's filed with MarketDragon's lip-sync tools, not the text-to-video pickers.

Seedance 2.5, 2.0, and 1.5 (ByteDance)

ByteDance launched Seedance 2.5 on July 31, 2026, its own write-up describing 30-second single-pass clips, up to 30 reference images, 10 reference video clips and 10 reference audio clips, and timestamp-level editing to fix one detail without re-rendering the whole clip (ByteDance Seed blog, source below). MarketDragon's config confirms the render-time numbers for bytedance/seedance-2.5-text-to-video: prompt cap 30,000 characters, duration 4-30 seconds, resolution 480p/720p/1080p, audio generated by default.

Seedance 2.0 (regular, Fast, and Mini) is the older, cheaper line — still active in MarketDragon. Regular caps duration at 15 seconds and adds a 1080p tier Fast and Mini don't offer; our config lists a 20,000-character prompt cap for all three, but that number is not yet verified against a live render — treat it as a ceiling, not a confirmed working length. Seedance 1.5 Pro is the oldest active tier: a 2,500-character prompt, up to two reference images, and audio that's off by default (generate_audio defaults false, unlike 2.5's default-on).

Hailuo 2.3 and MiniMax H3

MiniMax's own Hailuo 2.3 news page (source below) reports sharper micro-expressions, smoother full-body motion, and a wider stylization range — including anime — over the prior Hailuo 02. In MarketDragon, Hailuo 2.3 is image-to-video only: hailuo/2-3-image-to-video-pro and its Standard sibling take a starting image (no text-to-video mode), a 5,000-character prompt, 6 or 10-second duration, and 768P or 1080P resolution — though our own pricing notes flag that Kie doesn't publish a rate for the 10-second/1080P combination, so that pairing may not be reliably available.

MiniMax H3 is the newer, general-purpose successor. Coverage of MiniMax's own announcement (Hugging Face write-up, source below) reports H3 adding native stereo audio and first/last-frame control on top of Hailuo 2.3's motion work. In MarketDragon, minimax-h3/text-to-video and its image-to-video sibling are active: 7,000-character prompt, duration 4-15 seconds, resolution 768P or 2K. The image-to-video variant takes a first frame, a last frame, or both — no aspect ratio field, since the frames set it.

Side by side (MarketDragon config, not vendor marketing)

*Seedance 2.0's 20,000-character cap is the config ceiling, not confirmed on a live render — see above.

Which to pick for what

• Fights and fast action. Hailuo 2.3 and MiniMax H3 are reported to lead on complex body motion and camera work (MiniMax's own claims, above). Seedance 2.5's longer 30-second single pass also helps a fight that needs to land more than one beat without a cut.
• Dialogue scenes. Models with native audio and longer prompts give you room to write both action and line reads in one prompt: Seedance 2.5 (audio on by default), Wan 3.0/Prime (audio supported), and Gemini Omni Flash 1.1 (audio supported, plus first/last-frame control for reaction shots).
• Product shots. Wan 2.6 and Wan 2.7's single-image or first/last-frame inputs suit a still-to-motion product render more than a multi-reference model built for cast consistency. MiniMax H3's first/last-frame image mode works the same way for a bookended product turn.
• A consistent character across shots. Seedance 2.5's 30 reference images and Wan 3.0's 10 give the most room to carry a face, wardrobe, and location into every shot at once. See our companion guide on wiring a character node into every shot in Spaces for the workflow, not just the model limit.

Try them in Spaces

Every model above is in the Spaces model picker on a video node, and in the AI Director's prompt bar model selector. Planning a board costs nothing; a render is the only step that spends credits, and MarketDragon shows the cost before you press Generate. Check pricing for current credit costs.

Sources

• Google Developers Blog, Gemini Omni 1.1 Flash lets you build with more control, checked 2026-09-25 (official, dated August 27, 2026)
• Dataconomy, Alibaba Launches Wan3.0, A 30-second AI Video Generation Model, checked 2026-09-25 (reported)
• OpenRouter, Wan 3.0 Prime model listing, checked 2026-09-25 (reported)
• Kie.ai, Wan 2.7 Video API, checked 2026-09-25 (marketplace/vendor-adjacent, reported)
• Kie.ai, Wan 2.6 API, checked 2026-09-25 (marketplace/vendor-adjacent, reported)
• ByteDance Seed, Introducing Seedance 2.5, checked 2026-09-25 (official, dated July 31, 2026)
• MiniMax, MiniMax Hailuo 2.3 announcement, checked 2026-09-25 (official)
• Hugging Face blog, What Is MiniMax H3 (Hailuo 3.0)?, checked 2026-09-25 (third-party summary of MiniMax's own H3 announcement, reported)
• config/kie-models.php — MarketDragon's own model configuration, read 2026-09-25, for every MarketDragon-specific limit above (prompt caps, duration, resolution, reference counts, audio flags)
• Modules\AI\Models\AiModel (ai_models table, is_active) — checked 2026-09-25 via sail artisan tinker to confirm which of the above are actually available in MarketDragon today, not just present in config