How long AI clipping actually takes. Measured, not claimed.
Turning a podcast into vertical shorts with Skapo takes a median of 8.8 minutes to the first finished clip and 12.3 minutes to the whole set, producing a median of 19 clips per episode. Those figures come from 67 real production jobs and 2,329 rendered clips between 2026-04-19 and 2026-09-10, not from a marketing page.
We publish this because nobody else in this category does. Every figure below is a median with its tail, the sample size is stated, and the method is at the bottom. Reuse it with attribution.
Timing, measured from submit
| Metric | Value | What it means |
|---|---|---|
| First clip ready | 8.8 min | Median. Includes queue wait, download and analysis, not just render. |
| Full set ready | 12.3 min | Median. Every clip renders in parallel, so the set finishes shortly after the first. |
| Full set, slow tail | 24.8 min | 90th percentile. One job in ten is slower than this, usually a long or high-bitrate source. |
| Median source length | 50.4 min | The typical episode in this sample. Longer sources take proportionally longer to analyse. |
Output per episode
| Metric | Value | What it means |
|---|---|---|
| Clips per job | 19 | Median. The full range across the sample was 6 to 103. |
| Clips per source hour | 22.7 | Median, normalised for episode length so it compares across shows. |
| Clip length | 42s | Median. 80 percent land between 32 and 59 seconds, a consequence of cutting on complete thoughts rather than to a fixed length. |
| Job completion rate | 95.6% | Across 227 terminal jobs. Failures are mostly unreachable or malformed source files. |
Clips produced is not the same as clips worth posting. Expect to publish a subset; the ranking exists so the subset is the top of the list rather than a guess.
Method, and what these numbers are not
- Source:Skapo’s production database, every job between 2026-04-19 and 2026-09-10. Timing percentiles use the 67 jobs that produced at least one clip and carry a usable timestamp pair.
- Measured from submit,not from when a GPU worker picked the job up. Queue time is the user’s time, so it is counted. Measuring from render start would produce a faster number that nobody actually experiences.
- Medians with a tail, never a mean on its own. A mean render time is the figure most vendors quote and the one least likely to match your job.
- No per-language breakdown, deliberately. Most jobs in this window ran on automatic language detection, so a per-language table would present a claim about a few dozen jobs as if it were coverage. What each language actually supports is on the languages page.
- This is one tool’s data, not an industry benchmark. It says what Skapo does on real jobs. Other tools publish their own claims; where we compare, we date the claim and link the source.
Common questions
How long does it take to turn a podcast into shorts with AI?
Measured across 67 production jobs on Skapo, the median time from submitting a video to the first finished clip is 8.8 minutes, and 12.3 minutes to the complete set. The slow tail matters too: 10 percent of jobs take longer than 24.8 minutes, usually long or high-bitrate sources. Median source length in this sample was 50.4 minutes.
How many shorts can you get from one podcast episode?
The median job produces 19 clips, with a range of 6 to 103 across the sample. Normalised for length that is about 22.7 clips per hour of source video. Note that this counts clips produced, not clips worth posting: expect to publish a subset.
How long is an AI-generated short clip?
Median clip length is 42 seconds, with 80 percent of clips falling between 32 and 59 seconds. That distribution is a consequence of cutting on complete thoughts rather than to a fixed target length.
How reliable is AI video clipping in production?
95.6 percent of terminal jobs completed successfully across 227 jobs in this window. Failures are mostly unreachable or malformed source files rather than render errors.
Where does this data come from?
Skapo's own production database, covering every job between 2026-04-19 and 2026-09-10. Timings are measured from job submission, including any queue wait, not from when a GPU worker picked the job up. Figures are medians with a stated tail, and the page is regenerated twice a day.