Seedance 2.5, Wan 3.0, Omni Flash 1.1 and H3: four vendors, one wallet

Four launches, four doors

The vendor pages I read on 2026-10-05 point to different places for each model. ByteDance's Seedance 2.5 announcement says it is rolling out on Jimeng AI and Doubao Pro, with API access coming soon via BytePlus ModelArk. The Wan 3.0 README on GitHub links Wan Explore and QwenCloud. Google's Omni 1.1 Flash post lists AI Studio, the Gemini API, Google Flow and the Gemini app.

What one wallet changes

On Sume all four share the same shape. You call POST /v1/videos or POST /v1/video-router/generate with a bare catalog id, get a job id, poll or receive a webhook, and download a Sume-hosted artifact. Billing is the workspace USD balance: at submit Sume reserves provider list times 1.25, and usage.cost is the billable amount (Video generation docs).

That means one balance to top up, one GET /v1/balance to read, one ledger for reserved, captured and refunded rows, and one concurrency model across the four (Generation admission).

What it does not change

• Sume does not expose every vendor feature. Check each model's capabilities in GET /v1/video-router/models before assuming a vendor-page feature exists.
• Sume adds 25 percent to provider list, so a vendor's own console can be cheaper per second; you are paying for one integration and one wallet.
• Catalog limits differ: seedance-2.5 4 to 30 s, wan-3.0 2 to 30 s, minimax-h3 5 to 15 s, gemini-omni-flash-1.1 3 to 10 s.
• Top-ups are a dashboard operation; the public API reads balance and usage but does not create a top-up.

When to use a vendor directly

If you only ever use one model and a vendor offers a feature Sume does not list, such as a vendor-only editing mode, the direct route is the right one. If you compare models per brief, a single wallet and a single job API saves integration work.

Checking a vendor channel claim

Vendor availability moves quickly, so the table is dated 2026-10-05. The Seedance 2.5 page says the API is coming soon, which means that today the only API route named on that page is not live. The Wan 3.0 README names two destinations but gives no pricing. The Google post names four surfaces. When you plan a project around one of these, re-read the vendor page on the day you start.

What does not move is the shape of your own integration. With a single catalog and job API, you can write one client and change the model field.

A worked budget across four models

Suppose a pitch needs one 10-second clip from each model at 720p. Wan 3.0 is $1.25, Omni Flash is $1.25, Seedance 2.5 is about $5.78 and H3 at 768p is $0.75. The four clips are about $9.03 in one wallet, and you can read each captured amount from one ledger afterwards. The same exercise across four vendor accounts needs four balances and four invoices.

The table above is a map of vendor channels, not a recommendation. Which door is right depends on your volume and what you need. A team that needs an unreleased feature the day it ships will go to the vendor; a team that wants one balance, one API shape and one place to read spend will use the catalog. Neither choice changes the arithmetic: the rate is per second or per token, the duration is yours, and the total is the product.

If you are unsure, run the same prompt on two models through one client and compare the captured amounts. The ledger shows the real cost of each, which is a better basis for a vendor decision than a launch post.