Check a video request against the Sume catalog before the 400

Ported Sora requests fail in one predictable way: a value the new model does not list. Sume answers with 400 unsupported_capability and does not bill, so the fix is cheap, but a nightly batch that hits 200 of them wastes an hour. A pre-flight that reads GET /v1/videos/models and compares duration, resolution and aspect_ratio against the model's own lists catches them before submit.

OpenAI removed the Videos API and the sora-2 ids on 2026-09-24 and named no replacement, so every request you send now is a translation. The catalog is the contract to translate into.

What the catalog tells you

Each row in the data array carries supported_durations, supported_resolutions, supported_aspect_ratios, supported_frame_images, supported_input_references and a seed flag. Sume's docs state that the catalog controls validation: if the model does not advertise a value, the request gets a 400 and the API does not silently drop the field.

Two more rejections are model-independent. size and seed return 400 unsupported_parameter on every v1 model, so a request translated from a pixel size or a seeded call fails everywhere until you remove them.

The pre-flight script

The script loads the catalog once, then checks a request against any model id. The demo request is a typical 12-second, square, 1080p ask. Run it and each model prints either ok or the exact field and list that would have failed.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
r = requests.get("https://api.sume.com/v1/videos/models", headers=H, timeout=30)
r.raise_for_status()
CATALOG = {m["id"]: m for m in r.json()["data"]}

def problems(model_id, req):
    m = CATALOG[model_id]
    out = []
    for key in ("seed", "size"):
        if key in req:
            out.append(f"{key} is rejected with 400 on every model")
    checks = (("duration", "supported_durations"),
              ("resolution", "supported_resolutions"),
              ("aspect_ratio", "supported_aspect_ratios"))
    for field, listed in checks:
        if field in req and req[field] not in (m.get(listed) or []):
            out.append(f"{field}={req[field]!r} not in {m.get(listed)}")
    return out

old_style = {"duration": 12, "resolution": "1080p", "aspect_ratio": "1:1"}
for mid in CATALOG:
    print(mid, problems(mid, old_style) or "ok")

Reading the output

Against that demo request, gemini-omni-flash-1.1 should report problems on all three fields, because its window is 3 to 10 seconds and it accepts only 16:9 and 9:16. seedance-2 should print ok: 12 seconds is inside its 4 to 15 range, and 1080p and 1:1 are both listed for it. The script does not hard-code that; it reads the live lists, so it stays right when the catalog changes.

What the pre-flight cannot see

It does not check audio. Gemini Omni Flash 1.1 always produces synced audio and the API rejects generate_audio: false, but the catalog's generate_audio flag only says a model can make audio, not that it cannot be turned off. It also does not check reference limits such as at most 10 images and 3 short clips on Omni. Keep those in your adapter.

Treat a pre-flight failure as a routing decision, not an error: try the next model in your fallback list, or shorten the clip. The 400 from the API remains the final authority, so keep handling it.