Editing craft
What AI clipping tools get right, and the clips they miss
5 min read
AI clipping tools promise one thing: drop in an hour of video, get back a stack of short clips ready to post. Opus Clip, Vizard, Klap, Submagic and a handful of other clipping tools now sell that promise to podcasters, streamers and agencies. Two tests published this year show what the promise looks like in practice: fast, cheap per minute, and a first draft rather than a finished cut.
Both tests come from companies that sell a competing product, which matters for how to read them. We say who wrote what as we go.
What the tools actually do
The workflow is the same everywhere. The tool transcribes the source, scores stretches of it for how likely they are to hold attention, cuts the best-scored ones to a short length, reframes them to vertical and burns in captions. Some add a score per clip: Opus Clip's runs from 0 to 100 and is meant to tell you which clips to look at first.
That is a real saving. Finding thirty good seconds in a ninety-minute podcast is the slow part of the job, and a tool that hands you twenty candidates in a few minutes changes how the work feels.
How fast they are
Reap, which makes one of these tools, paid for nine of them in April 2026 and ran the same six videos through each: a bilingual podcast, a webinar, an interview, a talking head, a gaming stream and a sports reel. Its report measures the time from upload to the first usable clip on the ninety-minute podcast:
Transcription itself is close to solved. On English, the report puts every tool it measured at about 97 to 98% word accuracy against human captions, with other languages a few points lower. What separates the tools is mostly speed, language coverage, export limits and API access, not whether they can hear the words.
What gets thrown away
Speed is the easy half. The harder question is how many of the clips are worth posting. BIGVU, which sells its own auto-shorts feature, ran Opus Clip on two weeks of long-form content and published what it found in July 2026:
- clips generated over two weeks of content
- 50+
- of generated clips discarded
- 20–40%
- per minute of source, however many clips come out
- 1 credit
The article gives both "roughly 40%" and "20 to 40 percent" without reconciling them, and does not say how many source videos were used or how performance was measured. Read it as a direction, not a measurement.
The reasons it gives for the discards are the interesting part. The tool sometimes picks a moment that makes no sense without what came before it. Captions can drift out of sync. Clips are often good but not always publishable without someone checking them. On Opus Clip's cheapest paid plan, there is no editor at all, so whatever the AI cut is what you get.
Reap's report adds figures from public user reviews of Opus Clip rather than its own tests, and flags them as such:
- clip selection judged right, one speaker
- 60–70%
- judged right, several speakers
- ≈ 40%
- of clips needing manual cleanup
- ≈ 70%
The gap between one speaker and several is the pattern to remember. A monologue is easy to cut: the subject is always on screen and the sentence always belongs to them. A conversation is not. The punchline is one person's, the setup was the other's, and the camera framing that worked for the wide shot loses one of them in vertical.
Where that leaves the tools
None of this makes the tools useless. Both articles land in the same place: a first-pass tool, not a set-and-forget one. They are good at surfacing candidates, transcribing accurately and doing the mechanical work of resizing and captioning. They are weaker at the decisions that make a clip land: where it starts, what context it needs, and who should be in the frame.
The scores deserve the most caution. A virality score is a ranking of candidates, not a forecast, and the user reviews Reap cites call it unreliable. Use it to choose what to watch first, not what to post.
Our take
We build an editor, so we have a side here, and it is worth saying what it is. The failures both tests describe are the ones we have written about one at a time: picking the moment, which is a judgment about context and not a score (see what makes a moment clippable); keeping the right person in a vertical crop when two people are talking (why auto-reframe loses the subject); and trimming a clip by what is said rather than by guesswork (editing by transcript, not waveform).
Our reading of the numbers: automation has made finding candidates cheap, and that is a genuine gain. It has not made choosing cheap. If a fifth to two-fifths of the output goes in the bin, the time saved depends on how fast a person can review, fix and reject, which is an editing problem rather than a detection one. This section is ours.
Sources
- BIGVU, Jessica Becker, Opus Clip Tested 2026: Where the AI Wins (and the 40% You'll Discard), 2 July 2026: read the article. The discard rate, the credit model, the plan limits and the failure modes come from it. BIGVU sells a competing product, which the article itself promotes.
- Reap, The State of AI Video Clipping, 2026, 29 April 2026: read the report. The test corpus, the times to first clip and the caption accuracy figures come from it, as do the user-review figures it cites for Opus Clip. Reap sells one of the nine tools tested and ranks itself first.
- The last section is ours.