Cost per keeper AI video clip: retakes and failed jobs
One usable AI clip costs the clip price divided by your keep rate: at 25 percent, a $0.75 clip costs $3.00 per keeper. Includes a Python function.

Cost per keeper is the cost of a generated clip divided by the share of clips you keep. A 6-second Wan 3.0 clip at 720p costs $0.75; if you keep one in four, each usable clip costs $3.00. Failed jobs are not billed, but completed clips you reject are.
Formula and table
Billed cost per attempt equals the clip price times the share of attempts that complete. Cost per keeper is that figure divided by the keep rate.
| Model | Clip price | Keep rate 50% | Keep rate 25% | Keep rate 10% |
|---|---|---|---|---|
| Grok Imagine Video 1.5 720p | $0.075 | $0.15 | $0.30 | $0.75 |
| Wan 3.0 720p | $0.75 | $1.50 | $3.00 | $7.50 |
| Gemini Omni Flash 1.1 1080p | $1.125 | $2.25 | $4.50 | $11.25 |
| Kling 3 audio on 1080p | $1.26 | $2.52 | $5.04 | $12.60 |
Python
The function below includes an optional failure rate, because failed jobs release their reserve. Measure the keep rate from your own runs; the numbers in the table are inputs, not benchmarks.
def cost_per_keeper(clip_cost, keep_rate, fail_rate=0.0):
"""Failed jobs are not billed; completed clips you reject are."""
billed_per_attempt = clip_cost * (1 - fail_rate)
return billed_per_attempt / keep_rate
for name, cost in {"grok": 0.075, "wan720": 0.75}.items():
print(name, round(cost_per_keeper(cost, 0.25, 0.05), 3))Moving the keep rate
A keep rate is the cheapest lever you have. Doubling the keep rate halves the cost per keeper, which beats most price differences between rows.
- Draft at a low resolution, then regenerate only the picks.
- Reuse prompts that worked and note the ones that did not.
- Cap spend per run so a retry loop has a ceiling.
Worked example
A campaign needs 20 usable 6-second shots. At a 25 percent keep rate that is 80 attempts. On Grok at 720p, 80 attempts cost $6.00; on Wan 3.0 at 720p they cost $60.00; on Kling 3 audio on at 1080p they cost $100.80. If a quick draft pass lifts the keep rate to 50 percent, the Wan figure halves to $30.00.
The second-order effect matters too: each rejected clip still counts toward queue capacity while it runs, so a low keep rate also slows the batch.
Limits
Keep rate depends on prompt, model and taste, and the table shows illustrative inputs. The tables show the model amount. A workspace Agent Fee line can be added on top, so treat the job's own usage cost as the truth and check the live catalog before you budget; limits and rows change.
Sources
Related posts
More in Pricing
- Cost to render a 30-minute video on Sume: a $3.00 ceiling
Sume Timeline reserves $0.10 per output minute and caps a render at 1,800 seconds, so 30 minutes reserves at most $3.00. Length table and the free plan call.
- A daily 15-second AI video for 30 days: monthly cost
Posting one 15-second AI video a day for 30 days is 450 generated seconds: $5.63 on Grok Imagine Video 1.5, $56.25 on Wan 3.0 720p, $94.50 on Kling 3 audio on.
- ElevenLabs v4 after October 12: $0.08 per 1,000 characters vs Sume
ElevenLabs v4 returns to $0.08 per 1,000 characters after the October 12 promo, Turbo to $0.04. Sume TTS lists $0.0475 per 1,000. Script costs side by side.
- fal pricing units: per megapixel, per 1,000 tokens, per second
fal's price page quotes per second, per image, per megapixel, per 1,000 tokens and per 1,000 characters. Convert each to dollars per output first.
Written by Sume