Most writing about AI app builders is demos and opinions. We host the apps, so we can count. This is aggregate telemetry for every app created on VibeKit between June 16 and September 13, 2026, with internal and team accounts removed. Methodology is at the bottom, and you are welcome to quote any number here with a link back.
The short version
| What we measured | Result |
|---|---|
| Signups that came from an iPhone | 78% (web 20%, API 2%) |
| Apps that got an AI-built version deployed | 71% |
| Time from creating an app to the agent's first live deploy | 2.7 minutes median (middle half: 2.2 to 3.8) |
| Messages a typical app receives | 2 median, 19 at the 90th percentile |
| Model usage in a builder's first week | $0.64 median, $2.14 at the 90th percentile |
| Model calls behind one message | 7.5 on average |
Where people build: mostly from a phone
78% of the people who signed up in the window did it in our iPhone app, 20% on the web and 2% through the API. That skew is partly who finds us, but it is also a real shift: describing an app is a thing you can do with your thumbs, and a coding agent does not need you at a keyboard while it works.
How fast: under three minutes to something live
71% of apps got at least one version built and deployed by the agent, and the median time from tapping "create" to that first successful agent deploy was 2.7 minutes. A quarter landed in under 2.2 minutes; three quarters within 3.8.
That first version is small by design. Most people then iterate: deployed apps had a median of 2 successful deploys (90th percentile: 4), counting the starter.
What people build
The starting point people pick says a lot about intent:
| Starting point | Share of apps |
|---|---|
| Blank (described from scratch) | 41% |
| Landing page | 20% |
| Mobile-style app | 13% |
| Game | 10% |
| Chat assistant | 6% |
| Dashboard | 3% |
| Online store | 3% |
| SaaS starter | 2% |
| Portfolio and other | 4% |
"Blank" leads by a wide margin: when people can describe the thing directly, most skip templates entirely.
How much iteration an app gets
| Messages to the agent | Share of apps |
|---|---|
| None | 5% |
| One | 32% |
| 2 to 4 | 27% |
| 5 to 19 | 26% |
| 20 or more | 10% |
The median app gets two messages. One in ten becomes a real project with twenty or more. That long tail is where almost all of the interesting software comes from.
What it costs
For builders whose model usage we pay for directly (people who bring their own API key are not in these numbers), model usage in the first week after creating their first app came to:
| First week of model usage | Amount |
|---|---|
| Median | $0.64 |
| 75th percentile | $1.62 |
| 90th percentile | $2.14 |
| Average | $1.34 |
Two things shape those numbers. First, 97% of builders stayed within the $2.50 of free credit new accounts get, so the distribution is partly capped by what people are given; a small number of heavy builders spent far more. Second, one message to an agent is not one model call: the agent reads files, edits, runs checks and deploys, which averaged 7.5 model calls per message. Our earlier cost breakdown looks at total spend per app over a longer period.
What the first build does not tell you
Every number above is about the first session, because that is where building has become fast and cheap. The harder question for anyone choosing a tool is the second visit: whether your app is still live when you come back, what it costs to keep running after the free credit, and whether the agent still knows your project. Test that before you commit to a builder, not just the first five minutes.
Methodology
- Window: apps and signups created 2026-06-16 through 2026-09-13 (UTC). Data read on 2026-09-14.
- Exclusions: internal and team accounts. Deleted apps are included in template splits, message depth and deploy timing.
- Cost: metered model cost of platform-billed calls, before any markup. Bring-your-own-key calls go straight to the user's provider and are not visible here.
- Speed: time from app creation to the first successful deploy triggered by the agent (starter deploys excluded).
- Messages: messages a person sent to the app's agent.
- Percentages are rounded.