For a 5-minute window of a YouTube video, our pipeline takes about 7 minutes end to end and returns 5 clips. The part people call "the AI" — scoring every candidate moment — takes 2.5 seconds of that. Everything else is moving video around: downloading the source and rendering the output.
These are stage times from our own production worker on 20 August 2026, not estimates. One machine, one job at a time. The source was a 5-minute window of a YouTube video that was not already cached.
| Measured | Source | |
|---|---|---|
| Fetch the source (5-minute window) | 162 s | request a77e8c7d, ingest stage |
| Transcribe it | 137 s | request a77e8c7d, transcribe stage |
| Score every candidate moment | 2.5 s | request 53cecc91, highlight stage |
| Plan the crop | 19 s | request a77e8c7d, crop_plan stage |
| Render 5 clips | 82–215 s | requests 03a44092 and 7bce807c, render stage |
A 4-minute video downloads in about 3 seconds at 10 MB/s. A 20-minute window cut out of a 3-hour-49-minute source took 759 seconds at 58.8 KB/s — roughly 170 times slower per byte.
The reason is that pulling a section out of the middle of a long file is a different operation from downloading a short file, and it gets throttled. This is not specific to us: anything that reads a byte range out of a long video hits the same wall.
What that means in practice: if you are choosing a window, a shorter one is not just cheaper, it is disproportionately faster.
| Measured | Source | |
|---|---|---|
| 4-minute video, whole file | 3 s at 10.13 MB/s | measured in a container mirroring production, 20.08.2026 |
| 20-minute window from a 3h49m source | 759 s at 58.78 KB/s | same measurement run |
| 20-minute window from a 19-minute source | 589 s at 103.55 KB/s | same measurement run |
The category advertises clipping "in seconds". That is true of the scoring step and false of everything around it. We would rather you know which number is which before you plan a workflow around it.
We are not giving you a comparison against other tools here, because we have not run the same video through them. When we do, the page will say which video and which day.
No. Our pipeline returns up to 5 clips per run regardless of window length, so a longer window gives a better selection inside it, not more output. A 20-minute window costs roughly four times the download time of a 5-minute one for the same clip count.
Transcription processes every second of audio; scoring reads the finished transcript. On our hardware transcription took 137 seconds for a 5-minute window, and scoring took 2.5 seconds for a 3-minute one.
No. They are one machine, one job at a time, on a specific day. When several runs are queued together the per-stage numbers stretch a lot, because the time a job spends waiting counts as part of that stage.
5 free videos every month, no card required.