We Open-Sourced Our First 39 AI Video Generations - Here's the Data
August 5, 2026 · 6 min read
Most AI video tools publish marketing claims. We'd rather publish numbers. Here is the complete, unedited production data from Reeloop's first month live - every finished generation, with the method and the caveats included.
The dataset
Every video ever completed on Reeloop as of August 5, 2026:
- 39 finished videos (3 more were still rendering when we pulled the data)
- 12 distinct creators
- 4 video styles used (Viral Story, Educational, Documentary, UGC Ad) out of the 6 we offer
- 5 languages generated
- 38.4 seconds average video length
- 3.2 credits average per video (measured when a standard 30s video cost 2 credits; it is 3 today - see /pricing)
No cherry-picking: this is 100% of completed generations, not a highlight reel.
The denominator: 39 out of 54
Success rate matters more than the success count. Over the same window, 54 renders were started and 39 completed - a 72% success rate. The 15 failures were provider-side (a generation API stalling or refusing a scene), never silent: every failed render stopped the job, refunded the credits automatically, and emailed the creator. Nobody paid for a video they didn't get.
72% is not the number we want to run a business on, and we're publishing it anyway - because "it usually works" is the claim every competitor makes and nobody measures. The fixes since (per-scene retry, stuck-scene timeout detection, Director Mode review before render) target exactly this rate, and we'll publish the next cohort - with generation times - once the instrumentation we added for it has enough volume. That instrumentation is live now: the Proof & Data page carries rolling 7-day / 30-day / last-100 success rates recomputed daily from production, with generation-time percentiles publishing automatically once the measured sample passes 100 renders.
What people actually make
The style mix surprised us. The two most-chosen styles were Viral Story (8 videos) and Educational (6) - the formats built around a scripted narrative - while the two "spectacle" formats (ASMR, Italian Brainrot) weren't touched at all in month one. Documentary and UGC Ad picked up 2 each.
The early adopters aren't making memes. They're making content with a script - the exact use case a faceless channel is built on.
Average length of 38.4 seconds sits right in the retention sweet spot for Shorts and TikTok (under 60s), and the 3.2-credit average tells us most users generate standard videos with a few longer or premium ones mixed in - which is how we priced the tiers.
What an AI video actually costs
We track our own provider costs per video (AI footage, voiceover, stills, final render). The real blended cost of a 30-second video:
| Video type | Production cost |
|---|---|
| Standard 30s b-roll | ~$2.32 |
| 4K 30s | ~$7.72 |
| Talking-avatar 30s | ~$4.76 |
For comparison, a freelance short-form editor charges $25-150 per video. At 3 videos per week, that's $300-1,800/month against $35-79 for a Reeloop plan - the AI workflow is roughly 10-50x cheaper per finished video.
The honest limitations of this data
- 39 videos is a small sample. These numbers will move as volume grows - we'll publish an update at 500 videos.
- Generation time isn't in this study. Our current timestamps measure publishing, not render completion, so we won't quote a number we can't back. (We're instrumenting it properly.)
- Retention data belongs to the creators, not to us - so we can't yet correlate style with views. That's the study we most want to run with creator permission.
Why publish this at all
Because "AI video" is full of demos and short on production numbers. If you're evaluating whether an AI workflow can replace editing, you deserve real costs and real usage - not a landing page. This is the first of a series; the next edition covers generation success rate and time-to-video once our instrumentation catches up.
Calculate what your videos would cost →