Split a 181-second Reel into valid AI video jobs (even split)

A Reel can now run to 3 minutes, as NapoleonCat reports, and a 181-second cut is one second too long. The trap when you plan the jobs is the tail: six jobs of 30 seconds leave 1 second, and no video model on Sume takes 1 second (Wan 3.0 starts at 2). The fix is to pick the job count first and divide evenly.

The planner

Count jobs as ceil(total / max_seconds), then spread the seconds so each job is as close to equal as possible. Run it for your model's min and max.

def split(total, lo, hi):
    n = -(-total // hi)
    base, extra = divmod(total, n)
    lens = [base + 1] * extra + [base] * (n - extra)
    assert all(lo <= x <= hi for x in lens), lens
    return lens

for name, lo, hi in [("seedance-2.5", 4, 30), ("wan-3.0", 2, 30),
                     ("sume/auto", 3, 10), ("kling-3", 4, 15)]:
    lens = split(181, lo, hi)
    print(name, len(lens), lens)

Jobs and cost for 181 seconds at 720p

The limits come from the video catalog: Seedance 2.5 and Wan 3.0 reach 30 seconds, sume/auto (Gemini Omni Flash) takes 3 to 10, Kling 3 takes 4 to 15. Prices are the 720p per-second rates times the job lengths.

Then join them

Past 12 clips the final Timeline join auto-chunks, and a forced single render is refused above 12, which matters for the 19-job Omni plan. See the render_strategy post for the details. Totals exclude the Timeline render (3 minutes of output is $0.30, 181 seconds is $0.40).