Re-render your top old videos first under a fixed $3 budget

Rank every clip by monthly views per dollar of re-render cost, then take clips from the top until the budget runs out. With the Sume Omni 720p rate of $0.125 a second, a $3.00 budget re-renders four of the five sample clips below and skips promo-a, which does not fit.

OpenAI's deprecations page lists the Sora video models as removed on 2026-09-24 (read 2026-10-05). If your library was rendered there, you will not redo all of it, so the useful question is which clips earn a second render first.

Cost per clip is length times rate

The repo docs state that Sume bills the provider list price times 1.25 and rounds each job up to the cent. For gemini-omni-flash-1.1 at 720p the list price is $0.10 a second, so the billed rate is $0.125 and a clip costs seconds x 0.125, rounded up. Notice that length does the work: a three-second teaser with big traffic is the best deal on the list.

The views column is yours. Pull it from analytics, and use something stable like a 30-day count. The script only needs a number per clip.

The greedy fill

The script sorts by views per dollar and adds clips while they fit. Greedy is not optimal for every knapsack, but for a first pass it is easy to explain to whoever approves the spend, and the ranking column is what they will actually read.

With a $3.00 budget the teaser ($0.38), hero-intro ($1.00) and pdp-loop ($0.63) fit and sum to $2.01. Promo-a at $1.25 would take the total to $3.26 and is skipped, so promo-b at $0.75 fills the gap and ends at $2.76. Run the code to see the exact list, because the rule skips what does not fit and keeps going.

import math
RATE = 0.125  # gemini-omni-flash-1.1 at 720p: $0.10 list x 1.25
clips = [("hero-intro", 9100, 8), ("pdp-loop", 4200, 5), ("promo-a", 3900, 10),
         ("promo-b", 800, 6), ("teaser", 7600, 3)]   # (id, monthly views, seconds)
BUDGET = 3.00

def cost(seconds):
    return math.ceil(round(seconds * RATE, 6) * 100) / 100

spent, picked = 0.0, []
for cid, views, secs in sorted(clips, key=lambda c: c[1] / cost(c[2]), reverse=True):
    c = cost(secs)
    if spent + c <= BUDGET:
        spent += c
        picked.append((cid, c))
print(picked, f"total ${spent:.2f} of ${BUDGET:.2f}")

Before you spend it

Re-rendering is a new creative act. Omni renders native audio, so the new clip will sound different from the old one, and a prompt tuned for the old model may need edits. Render the single top clip first, compare it to the original, and only then release the rest of the budget.

Submit each clip with an idempotency key built from the clip id and the model, so a restarted batch cannot bill a clip twice.

• Use 720p for the first pass. 1080p on Omni is $0.1875 a second, 50 percent more.
• Keep the old file. A re-render replaces nothing unless you delete it.
• Record the billed cost from each finished job and compare it to this estimate.

Revisit the ranking monthly

Views per dollar is a snapshot. A clip that was cold in September can spike in October, and the budget you set aside this month should follow the traffic. Keep the clip table in a file, refresh the view counts on a schedule, and rerun the script against whatever budget remains.

Record each clip you re-render, so the next pass does not rank it again. A one-column list of finished ids is enough.