Which new video model for which job: Seedance, Wan, Kling, H3, Omni

For clips up to 30 seconds with references, use seedance-2.5; for the lowest price per second, wan-3.0; for a 15-second clip with a sound toggle, kling-3; for built-in stereo sound with 2K or 4K upscales, minimax-h3; for 3 to 10 seconds at 4K, gemini-omni-flash-1.1. Veo 3.1 and LTX-2.5 are not in the Sume catalog. These are the facts as of 2026-10-06.

Vendor dates are from Magic Hour's tracker. Limits and prices are from the Video Router docs and video docs.

The decision table

Prices are billed on Sume per second, at the lowest and highest listed resolution. Seedance is per token, so its range uses the 16:9 frame.

By job

Test prompts and drafts: Wan 3.0 at 480p, from 2 seconds. Brand clip that must hold a product or a face: Seedance 2.5, which takes image, video and audio references plus first and last frames. A 15-second ad with optional sound: Kling 3. A clip with soundtrack built in: MiniMax H3, which can also be lip-synced with H3 Max. A short high-resolution clip: Omni Flash, which is the only one of the five to reach 4K natively and is limited to 16:9 and 9:16.

References are where the models differ most. Wan 3.0 accepts up to 10 reference images, 5 reference videos and 5 reference audio clips. Gemini Omni Flash 1.1 accepts up to 10 images and 3 videos of at most 3 seconds each, with no audio reference. Kling 3 on Sume takes text-to-video or first and last frames only. If your brief is a product photo or a brand voice, that list decides the model before price does.

Switching is one field

Every model uses the same request; only model and the values it allows change. This loop tests all five with one prompt at 5 seconds, skipping nothing the catalog would refuse.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
base = {"prompt": "A chef plates a dessert, overhead shot", "duration": 5,
        "resolution": "720p", "aspect_ratio": "16:9"}
for m in ("wan-3.0", "seedance-2.5", "kling-3", "gemini-omni-flash-1.1"):
    r = requests.post("https://api.sume.com/v1/videos", headers=H,
                      json={**base, "model": m}, timeout=60)
    print(m, r.status_code, r.json().get("id"))
# minimax-h3 has no 720p; use 768p
r = requests.post("https://api.sume.com/v1/videos", headers=H,
                  json={**base, "model": "minimax-h3", "resolution": "768p"}, timeout=60)
print("minimax-h3", r.status_code, r.json().get("id"))

A 5-second 720p test across the first four costs $4.85 with Kling's sound off, plus $0.38 for H3. Read the results, then confirm values before scaling with the catalog check. If you came from Veo, this note lists what Sume offers instead.

What is not on this list

Veo 3.1 (2025-10-15) and LTX-2.5 (2026-08-11) are dated on the same tracker but are not in Sume's catalog, so you cannot call them through Sume. Kling 3 on Sume is the 3.0 generation listed as kling-3, with no reference inputs. MiniMax H3 Max, a sibling of H3, adds 1080p and is the one with the lip-sync endpoint. Seedance 2.0 and its Fast and Mini variants remain listed at up to 15 seconds.

The table is a snapshot. Read /v1/videos/models for the live list before you build, and treat any model not shown there as unavailable.

If you only run one test this week, run the same 5-second prompt on Wan 3.0 and Seedance 2.5 at 720p and compare. The two cost $0.63 and $2.89, and the difference you see tells you whether the premium for references and first and last frames is worth paying for your work.

Then pin the winner's id, resolution and length in config, and revisit the choice when the catalog changes. Models in this class are updated often, so re-run the same prompt when the catalog changes.