Price a 20-prompt video regression suite: 360p, 3 seconds on Sume

Twenty golden prompts at 3 seconds and 360p on gemini-omni-flash-1.1 cost $0.12 each on Sume, so a full run is 20 x 0.12 = $2.40. That is cheap enough to run on every prompt-template change, which is the point of having a regression suite at all.

OpenAI removed the Sora video models on 2026-09-24 per its deprecations page. After a change like that, the only defense against the next one is a set of prompts you can rerun against whatever model you adopt, with a known price per pass.

The arithmetic

The repo docs list Omni at $0.03 a second at 360p before margin. Sume bills list times 1.25, so 3 seconds is 3 x 0.03 x 1.25 = $0.1125. Each job is rounded up to the cent, so one prompt bills $0.12. The rounding happens per job, which means 20 jobs cost 20 x 0.12 = $2.40, not 20 x 0.1125 = $2.25.

At 720p the same three seconds is 3 x 0.10 x 1.25 = $0.375, billed $0.38, and 20 prompts cost $7.60. So 360p is the right tier for catching a broken prompt and 720p is for the final look.

The cost check

The script keeps a three-row suite inline so it runs as written; load your own from a JSONL file with one object per line. It prints the cost of a run at both tiers and the exact request body for the first prompt. Compare the printed total to what the dashboard reports after a real run, and investigate if they differ.

import json, math
RATE = {"360p": 0.0375, "720p": 0.125}   # Omni per second: list x 1.25
SUITE = """\
{"id": "mug", "prompt": "Steam rising from a ceramic mug, slow dolly-in"}
{"id": "rain", "prompt": "Rain on a window, city lights blurred behind"}
{"id": "shoe", "prompt": "A white sneaker turning on a pedestal, soft light"}
"""
def job_cost(seconds, res):
    return math.ceil(round(seconds * RATE[res], 6) * 100) / 100

def body(row, res="360p", seconds=3):
    return {"model": "gemini-omni-flash-1.1", "prompt": row["prompt"],
            "duration": seconds, "resolution": res, "aspect_ratio": "16:9"}

rows = [json.loads(line) for line in SUITE.splitlines()]
for res in ("360p", "720p"):
    total = sum(job_cost(3, res) for _ in rows)
    print(f"{len(rows)} prompts, 3 s at {res}: ${total:.2f}")
print(json.dumps(body(rows[0])))

What makes the suite useful

Keep prompts that failed in the past. A clip that once rendered the wrong object, or produced a face you disliked, belongs in the set permanently. Add one prompt per feature you depend on, such as a reference tag, a vertical aspect or a first frame.

Render the suite after every change to a prompt template, and keep the outputs in a folder named for the run. Looking at twenty three-second clips takes a few minutes and catches regressions that no assertion can.

• Do not compare clips for equality; Sume's /v1/videos rejects seed, so renders vary.
• Use the same idempotency key prefix per run label, so a rerun of a label never double-bills.
• Pin the model id in the suite file and change it deliberately.

When the price changes

The rates in the script are the repo-documented list prices times 1.25 as read on 2026-10-05. If the pricing docs change, update the two numbers and the totals follow. The arithmetic stays the same: seconds times rate, rounded up to the cent per job, then multiplied by the number of prompts.

A suite is also a cheap way to see whether a different model is worth switching to. Render the same twenty prompts on two ids at their lowest shared resolution and compare the costs and the clips side by side.