Clamp video duration when swapping Sora for a Sume video model

After the Sora API shutdown, the length your code asked for may not fit the Sume model you swapped to, because each id has its own duration range. Clamp the requested seconds to the range of the id you are about to call. OpenAI's deprecations page lists Sora 2 and the Videos API as shut down September 24, 2026, with no replacement named, so the model choice is yours.

These ranges come from Sume's video model docs and catalog, read 2026-10-06.

Duration ranges per Sume video id

A clamp function

Log every clamp. A 12 second request that becomes 10 seconds changes both the clip and the price, and you want that visible in your own logs.

LIMITS = {
    "seedance-2.5": (4, 30),
    "seedance-2": (4, 15),
    "seedance-2-fast": (4, 15),
    "seedance-2-mini": (4, 15),
    "wan-3.0": (2, 30),
    "kling-3": (4, 15),
    "minimax-h3": (5, 15),
    "gemini-omni-flash-1.1": (3, 10),
    "h3-max-recast": (5, 30),
}

def clamp_seconds(model, seconds):
    low, high = LIMITS[model]
    fixed = max(low, min(high, seconds))
    if fixed != seconds:
        print(f"clamped {model}: {seconds}s -> {fixed}s")
    return fixed

if __name__ == "__main__":
    print(clamp_seconds("gemini-omni-flash-1.1", 12))
    print(clamp_seconds("seedance-2.5", 12))

Other fields that differ

Duration is only one of the fields. Sume's video models have no seed, reject size, and return 400 for a non-empty provider.options. Resolution lists differ too: kling-3 takes 720p and 1080p, and gemini-omni-flash-1.1 takes 360p up to 4K with native audio always on. Start from the Sora port guide and pin the id with the canonical slug field.