Which Video Model Should You Use? Seedance, Kling, Hailuo, and Wan by Task

There is no sensible “best video model” default. A five-second product animation, a reference-driven character shot, a multi-shot story, and a repair of an existing clip fail in different ways. Treat them as different production jobs.

The 2026-07-20 gateway-reconciled catalog currently covers the four core families in our routing brief: Seedance, Kling, Hailuo, and Wan. The useful public conclusion is not that one family wins. It is that a cost-controlled workflow should use a lower-cost candidate pass where appropriate, then reserve premium or reference-heavy work for the shots that survive selection.

The short answer

• Seedance is the starting point for a story-led or complex concept where the relationship between action, shots, and audio matters more than a fixed source asset.
• Kling is the upgrade path for a high-value creative or movement-heavy shot, especially when the selected concept needs more visual punch rather than more cheap retries.
• Hailuo is the practical first pass for image-to-video, people, character, or motion experiments where speed matters.
• Wan is the default when the job is a controlled production operation: text-to-video, image-to-video, reference-video work, or editing an existing clip are separate tools, not interchangeable names.

The position is deliberately opinionated: do not send every request to a flagship model, and do not send a reference-video or edit job to a generic text-to-video route. The wrong operation type wastes more money than a small price difference between models.

What each family is for

Cost is more than seconds and resolution

The current catalog makes the difference visible. Hailuo 2.3 Fast is listed at $0.1875 and Hailuo 2.3 at $0.2778 in their representative configurations. Wan's text-to-video and image-to-video routes use a displayed per-second unit of $0.1, while reference-video and editing routes must be evaluated with their own input and output rules: $0.35 and $0.2 respectively.

Those are useful starting inputs, not a promise that every task has the same total. The actual production cost is:

usable-shot cost =
  (all attempts + input-video seconds + output seconds + repair time) / accepted shots

This is why “cheapest per second” is not a routing rule. A short, low-cost draft that needs to be remade three times can lose to a more reliable final call. Likewise, a reference-video task can look expensive only because its input footage is part of the work; comparing it to a blank text-to-video call is comparing different products.

A repeatable two-stage production method

The routing brief's most useful operational recommendation is a two-stage workflow. It is reproducible without claiming an unpublished quality benchmark.

1. Write a shot card: source assets, action, desired continuity, duration, resolution, audio need, and a maximum budget.
2. Generate a limited number of candidates with a lower-cost or faster route when the task permits it.
3. Select a candidate against the one criterion that matters most for the shot—product shape, character stability, movement, story beat, or edit fidelity.
4. Upgrade only selected candidates to Seedance Standard, Kling, or a higher-control Wan operation.
5. Record attempts, accepted shots, input/output seconds, actual billed cost, and repair time.

Use this record for every candidate family:

{
  "task_id": "product-bottle-motion-01",
  "operation": "image_to_video",
  "source_assets": 1,
  "duration_seconds": 5,
  "resolution": "720p",
  "model": "wan2.7-i2v",
  "attempts": 3,
  "accepted": 1,
  "primary_check": "product shape preserved",
  "actual_cost_usd": "record actual billed amount",
  "repair_minutes": 0
}

The internal test plan contains 40 fixed tasks: product image-to-video, people or character motion, ad storyboards, and short narrative scenes. It calls for repeated runs, P50/P95 latency, acceptance rate, total retries, human quality review, and usable-shot cost. It has not yet produced a public benchmark, so this article does not attach invented scores to the four families.

Rights checks belong before model selection

“Can accept a reference image or video” does not mean “may use any person's face, voice, or source clip.” Make the following fields part of a video job before routing it:

asset type → person or rights owner → consent status → allowed use → approved model path

For a fictional character, standard safety review may be sufficient. For a customer, employee, contracted talent, public figure, product asset, voice, or reference video, retain explicit rights and intended-use information. Do not route unverified real-person or copyrighted material automatically, and do not attempt to work around upstream safety filters.

Where this guide is deliberately limited

This is a current-catalog routing guide, not a promise about provider availability, content-policy outcomes, queue time, or universal video quality. Seedance and Kling configurations vary by upstream channel; the live pricing page is the authoritative display for the configuration you can call. Hailuo and Wan prices still need to be interpreted alongside duration, resolution, input material, and retries.

The next useful publication is not a louder ranking. It is a benchmark report with a fixed task set, repeated runs, acceptance criteria, and raw methodology. Until then, start with the operation type, measure usable-shot cost, and let selected scenes—not vendor branding—earn the premium path.