Skip to content
VodLift

Our numbers, and the ones we missed

We test VodLift before we claim anything, and we publish the method. Tested 12 Nov 2026 on Apple M4 Max, development build.

Four test streams. Speed numbers come from one Mac, in a development build, except the rows marked as the live site. They show what the pipeline does today. They are not a promise for the launch service.

The test streams

Public VODs, used for internal tests only. We do not show them here and we do not name the streamers.

Talk and news, 1 h 32 min, full camera and screen share
5 clips. Most numbers on this page come from this stream. API cost $0.045, or $0.029 for each stream hour.
Shooter, 2 h 9 min
5 clips. API cost $0.042, or $0.019 for each stream hour.
Strategy game, first 2 hours
5 clips. API cost $0.035, or $0.017 for each stream hour.
Quiet streamer, survival game, 1 h 50 min
5 clips.

Speed

Stream length 92 minutes.

Speed
Speech-to-text213 s26 times real time
Audio analysis4.8 s
AI scoring and selection29 s22 API calls
Render of 5 clips, 1080x1920 at 60 fps59 s217 s on a busy Mac
Whole run, upload not included5 to 8 minutes

On the live site

The site runs on a server. The processing worker runs on a Mac and takes jobs from that server. This is a closed test.

On the live site
Upload of 105 secondsClip 30 s after the upload ended
A 7-minute file, job start to clips on the server95 s

Cost to us

What the Claude API cost in our tests. This is our cost, not your price.

Cost to us
Each broadcast in our tests$0.008 to $0.045
Each stream hour, on the long streams$0.017 to $0.029Three long streams

Export quality

Measured on the rendered MP4 files.

Export quality
Resolution and frame rate1080x1920 at 60 fps
Video bitrateAbout 6.5 Mbps, H.264
Audio loudness-14.8 and -13.9 LUFSAAC 192 kbps
Audio and video syncWithin 16 ms

Captions and layout

Small samples. We say so in each row.

Captions and layout
Caption timing, median0.01 s171 word starts, 20 minutes of audio. Before the fix: 0.14 s
Caption timing, mean0.055 sBefore the fix: 0.155 s. Our target is under 0.1 s
Lit word matches the spoken word8 of 8 framesSmall sample
Camera on top, content below, no black bars2 of 2 clipsScreen-share clips. Small sample
Face in frame after face tracking10 of 10 samplesOne 42-second clip. Small sample

One run against Opus Clip

The same 92-minute talk stream, put through Opus Clip on 12 Nov 2026. A trial account, default settings (Clip long video, clip length Auto). It took the Twitch link. The first 3 silent minutes were cut to fit the trial, so 88 of 90 credits were used.

One run against Opus Clip
Time from start to clipsAbout 5 to 8 minutes on our Mac, file already on diskAbout 24 minutes, link import included
Clips made5, by design44
Clip length28 to 40 s26 s to 2 min 58 s in the first 18 clips. 12 of those 18 are longer than 60 s
Score shown50 to 66 of 100, with three written reasons for each clip92 to 99 virality score for the first 18 clips. No reasons shown in the list
What the picks areShort moments with a hook and a payoff: a joke, a rant, an argument with chatTopic summaries of the news story: 11 of the first 18 titles are about the same company story
Cost to the user for this stream1.5 of 15 hours on Starter ($15)88 of 150 credits on Starter ($15)

What this shows, and what it does not

  • This is one run on one stream. It is not a test of speed, and it is not a ranking.
  • Our time is from a development build on a fast Mac and does not include an upload. This is not a fair cloud-to-cloud comparison yet.
  • The picks are different in style: Opus Clip cut the stream into topic explainers, we picked short moments. For a streamer who wants funny-moment clips our list fits better. For a news channel the Opus list has value.
  • No person has rated the clips yet, so we do not say that ours are better. That is not proven.
  • Opus Clip showed a score of 92 to 99 on all 18 clips we looked at, which does not help a user choose. That is our reading, not a measurement.

Not compared yet: caption accuracy, layout quality (the Opus thumbnails did not load in the list), and the time position of each Opus clip.

What failed

Two results that missed our own targets.

Caption timing and face framing
Both missed our targets on the first test: captions were 0.14 s late (median) and the face left the frame when the streamer leaned. We fixed both and measured again. The new numbers are above, on small samples.
A scene change inside one clip
Not handled. A clip that crosses two scenes can show the wrong layout. Planned.

Not measured yet

Until these are done, we make no claim about them.

  • Pick quality. A blind test where people rate whether they would post each clip, on five streams. This is the test that decides if the picks are good, and we have not run it.
  • Overlap with the clips your viewers made on public VODs.
  • Other competitors. We have run only the one Opus Clip test above. We show no numbers for Eklipse or StreamLadder.
  • Speed in the cloud. The processing worker runs on a Mac, so our numbers are not cloud numbers.
  • Word error rate of the captions, on 30 clips.
  • How often a job ends with a downloadable clip.

We add each result here when we have it, with its method. Prices and features of other tools are on the pricing page.

Get in before it opens

VodLift is in a closed test and not open to the public. Leave your email and we write once, when the first spots are ready.

One email when it opens. We tell nobody your address.