The duration window all fallback video models share: 5 to 10 s on Sume

If you keep a fallback video model so a feature survives an outage, send a duration inside the overlap of every model in the chain. For gemini-omni-flash-1.1 (3 to 10 s), the Seedance 2.0 tiers (4 to 15 s), wan-3.0 (2 to 30 s) and minimax-h3 (5 to 15 s) the overlap is 5 to 10 seconds. A 12-second request would fail on Omni, and a 4-second one would fail on H3.

The Sora API had no successor named by OpenAI (Pondero), which is the lesson: pick a window that a second backend can also serve.

Overlap by chain

Intersections use the per-model ranges stated in the Sume docs. A chain's safe window is the highest minimum and the lowest maximum.

Why the low end matters more than the high end

Most prompts are written for a clip of a given feel and length. Raising the minimum to 5 seconds costs you very little for a product shot, and a lot for a 3-second loop. If your clips are shorter than 5 seconds, keep Omni, Wan and a model with a lower floor in the chain, and skip H3. If your clips are longer than 10, drop Omni.

Enforcing it

Do not hope the request fits. Compute the window from the live catalog at startup: take the max of every model's smallest supported_durations entry and the min of every model's largest. Reject a feature configuration whose duration falls outside it. The Node check shows the catalog read.

For the retry itself, keep one Idempotency-Key per model attempt. A replay on the same model returns the original job; a different model is a different request and needs a different key.

Other fields do not overlap either

Resolution and ratio narrow in the same way. Omni takes only 16:9 and 9:16. H3 renders native 480p and 768p. A chain with an Omni entry should use one of those two ratios; a chain with H3 should use 480p or 768p, not 720p, which the docs say H3 does not have. Audio is similar: Omni always has it, so a fallback to a silent-capable model changes the soundtrack of your output.

How to use the shared window

If your fallback chain is built from models whose ranges overlap, pick durations only inside the overlap. Nothing in that case needs a clamp at request time, and a failover from one model to the next never changes the length of the clip or fails validation. Set the app's own limit to the overlap and keep the model catalog read as a guard, so a catalog change that narrows a range shows up in a test rather than as a failed job.

Outside the overlap, you need an explicit rule. The simplest is to refuse the request. The next simplest is to split a long clip into several jobs inside the window and join them afterward, though that changes continuity between shots, so check the result by eye.

What it costs you to stay inside

A window shared by every model in the chain is the narrowest of the individual windows, so you give up the long clips that only some of them can make. That is a product decision: long clips on wan-3.0 or seedance-2.5 (up to 30 seconds each, per the Sume docs) can run on a separate path with its own fallback.