Best AI tool to turn long videos into shorts

Round 1: free tiers · tested Oct 18, 2026

We uploaded the same 37:50 interview to OpusClip, Vizard and Klap through each tool's free entry point and measured the top 3 clips each tool returned. OpusClip's free entry point turned out to be a 7-day trial of its paid Essential plan, not a free plan. Klap's free trial made clips but would not let us download them. CapCut, Descript and Submagic were not tested in this round.

Scores at a glance

Scores use the published 100-point rubric. 'Not scored' means we did not measure it; it does not mean zero. Totals are shown out of the points we could measure, so each one is partial.

OpusClip

Partial: trial of a paid plan, not a free tier

Plan used: 7-day Essential plan trial (started automatically, no card). Partial. Clip quality and cost are not scored, and export expiry was not checked.

Criterion	Score	Evidence
Clip quality max 30	Not scored	Needs two blind human raters, which we haven't run yet. Boundary check against the reference transcript: all 3 clips start and end on a complete thought (clip 1 drops a leading 'Well,'). Clip 3 is entirely contained in clip 2. Clip 2 opens with a 6 s hook quote taken from later in the same answer. OpusClip also removed one restart inside clip 1. Lengths are 1:08, 1:57 and 0:43; we asked for 30–60 s.
Caption accuracy max 20	19.9	3 errors in 658 reference words (WER 0.5%): 'guaranteed' for 'guarantee', 'where' for 'we're', and one dropped 'that'. No clip has a median caption delay over 0.5 s (median +0.3 s), so no timing deduction.
Speaker framing max 15	13.8	The active speaker's whole face is in frame in 92.1% of 229 frames (sampled 1 per second). Misses are cutaways to the listener, mostly reaction shots that are already in the source edit.
Time max 10	8 (provisional)	21–26 min from upload start to clips ready (upload ~4 min, processing 17–21 min); we scored the slow end. Round B fix time is not included yet.
Cost max 15	Not scored	Rubric cost is per postable clip on the cheapest plan that exports watermark-free 1080p. We did not test a paid plan. Free-tier result: $0, using 37 of the trial's 90 credits.
Watermark & export max 10	4 / 8 measured	Watermark: 0 of 4 ('OpusClip' logo on the downloads). Resolution: 4 of 4 (1080×1920 MP4). Export expiry: not checked (2 points unscored).

Run details: times, credits, limits

Upload	~4 min
Processing	~17–21 min after submit
Total	~21–26 min
Credits used	37 of 90 (trial)
Clips returned	33, 0:26–2:16, virality scores 74–99
Top clip	'Designer Babies: Should We Choose Our Child's Traits?' (score 99, 1:08)
Captions	Animated word-by-word captions with colored keywords; hook headline on the top 10 clips
Framing	Active-speaker face tracking
Export	1080×1920 H.264 MP4, 'OpusClip' watermark

12-second previews (360p, with audio)

Vizard

Partial: free plan

Plan used: Free plan (60 credits/month, 1 GB / 1 h upload cap). Partial. Clip quality and cost are not scored, and export expiry was not checked.

Criterion	Score	Evidence
Clip quality max 30	Not scored	Needs two blind human raters, which we haven't run yet. Boundary check: clips 1 and 3 start and end on complete thoughts. Clip 2 ends mid-sentence ('…as humans are multigenic' and the speaker continues 'meaning there are…'). Lengths are 0:47, 0:35 and 0:50, all within the 30–60 s target.
Caption accuracy max 20	19.9	2 errors in 324 reference words (WER 0.6%): 'is' for 'it's', and the last word ('faster') is missing because the outro card covers it. Timing is tighter than OpusClip's: median caption onset +0.1 s.
Speaker framing max 15	13.5	The active speaker's whole face is in frame in 89.7% of 126 frames (1 per second; the end card is excluded). In clip 2 the camera stays on the listener for 8 s while Dan Rather asks his question. Split-screen frames showing both faces count as in frame.
Time max 10	10 (provisional)	17 min or less from upload start to clips ready (upload ~6 min, processing 10 min or less). Round B fix time is not included yet.
Cost max 15	Not scored	We did not test a paid plan. Free-tier result: $0, using 37 of 60 monthly credits.
Watermark & export max 10	0 / 8 measured	Watermark: 0 of 4 ('Made with vizard.ai' on the video plus a branded end card). Resolution: 0 of 4 (720×1280 maximum on Free). Export expiry: not checked (2 points unscored).

Run details: times, credits, limits

Upload	~6 min
Processing	10 min or less
Total	17 min or less
Credits used	37 of 60 (monthly free allowance)
Clips returned	10, 0:33–1:06, scores 9.2–10.0
Captions	Word-by-word captions with colored keywords and emojis; title caption
Framing	Face-centered crop plus a two-speaker split screen; continuous tracking not confirmed
Export	720×1280 H.264 MP4, 'Made with vizard.ai' watermark, ~2 s branded end card

12-second previews (360p, with audio)

Frames and previews are excerpts of each tool's output from "Conversations in Science with Dan Rather & Jennifer Doudna – CRISPR" by Wonder Collaborative, licensed CC BY 3.0, clipped, reframed and captioned by the named tool and cropped and compressed by BenchedTools. Dan Rather, Jennifer Doudna and Wonder Collaborative are not affiliated with BenchedTools and do not endorse any tool or this ranking. Tool watermarks appear as exported.

What we measured, and how

• Caption accuracy. We OCR'd burned-in captions at 5–10 frames per second with Tesseract 5.5. We compared them with our corrected reference transcript after removing fillers and restarts. Every disagreement was then checked by eye on the frames, and OCR misreads were discarded. Both tools' raw OCR error rates before review were 10–35%, almost all of it OCR noise.
• Caption timing: when each caption appeared on screen compared with when the word starts in a Whisper transcription of the clip's own audio.
• Clip boundaries: we transcribed each clip's audio with Whisper and located it in our reference transcript, which shows whether a clip starts or ends mid-sentence and whether the tool cut words out.
• Speaker framing: 1 frame per second, scored by one reviewer (an agent, not blind) against who is speaking in the reference transcript.
• Resolution, codec and watermark, read from the downloaded files with ffprobe and checked on the frames.

What we did not measure

• Clip quality (30 points). The rubric needs two blind human raters; we have not run that.
• Cost (15 points). We tested no paid plans, so cost per postable clip on the cheapest watermark-free 1080p plan is unknown.
• Export expiry (2 points). We did not check how long downloads stay available.
• Round B (up to 10 minutes of manual fixes) and its effect on time. Only Round A defaults were measured, and only 3 clips per tool instead of 5.
• Klap: nothing that depends on an exported clip. CapCut was skipped because the email is already tied to an existing account. Descript sign-up was blocked by bot protection. Submagic needs a paid add-on.

Coming next

• Blind clip-quality ratings by two human raters for each tool's top 5 clips.
• Paid-plan runs (only with the owner's approval) to score cost and watermark-free 1080p export for OpusClip, Vizard and Klap.
• Round B timing, export expiry checks, and clips 4–5 for each tool.
• CapCut, Descript and Submagic, once account access is sorted out.

Data

• Caption error list and per-clip numbers (JSON)
• Framing judgments, one row per sampled frame (CSV)
• Full method and scripts (Markdown)