OpenRouter output_modalities=video vs Sume /v1/videos/models

OpenRouter documents three ways to find its video models: a dedicated GET https://openrouter.ai/api/v1/videos/models endpoint, the standard Models API with output_modalities=video, and a filter on its web models page. Sume follows the first one: GET https://api.sume.com/v1/videos/models returns a data array with the same capability fields, and GET https://api.sume.com/v1/catalog is the wider public model catalog. If your code already lists OpenRouter video models, the change is the base path and key.

Both facts were read on 2026-10-01 from OpenRouter's video generation guide and Sume's Video Generation docs.

What is the same and what is different?

Sume's docs state that everything on the /v1/videos route matches the OpenRouter Video Generation API except a short list of deltas. For discovery, the deltas are the base path, the key, and the id format.

Which fields should you read before submitting?

Each Sume model entry lists supported_resolutions, supported_aspect_ratios, supported_sizes, supported_durations, supported_frame_images, supported_input_references, generate_audio, seed, and pricing_skus. Limits are not uniform, and the docs call out several: seedance-2.5 takes 4 to 30 seconds, wan-3.0 takes 2 to 30, gemini-omni-flash-1.1 takes 3 to 10, and every other catalog model tops out at 15 seconds. Reading the list is cheaper than guessing and getting a 400.

import os, requests

r = requests.get(
    "https://api.sume.com/v1/videos/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
for m in r.json()["data"]:
    d = m["supported_durations"]
    print(m["id"], f"{min(d)}-{max(d)}s", m["supported_resolutions"], "seed" if m["seed"] else "no-seed")

How should you wire discovery into a pipeline?

Fetch the list at startup or on a short cache, then validate each request against the entry for the model you chose: is the duration in supported_durations, is the resolution in supported_resolutions, does the model accept the reference type you plan to send. Sume's docs say size returns 400 unsupported_parameter because every v1 model reports supported_sizes: null, and a seed is rejected because each model reports seed: false. Both are visible in the list before you spend anything.

If you also use OpenRouter, keep the two catalogs separate in your code. Ids differ (org/slug against bare ids), and a model that exists on one list may not exist on the other. Map ids explicitly rather than stripping prefixes.

Limits

I did not verify that Sume's general model list accepts an output_modalities filter, so the table says it is not documented. The live list changes as models ship or retire, so do not hard-code it; h3-max-recast and higgsfield-genjutsu are described as listed only when their provider is configured.