How long AI website generation takes, from recorded jobs

AI website generation takes from about 3 minutes to about a quarter of an hour, depending mostly on how much checking happens before the site is called done. In the 13 completed jobs recorded here, single agent website runs took 2.6 to 2.9 minutes, and the multi step pipeline, which plans, builds in steps, reviews screenshots and fixes, took 4.8 to 13.6 minutes, with one outlier at 30 minutes that included a long queue.

Where the numbers come from

The table comes from the job records of OneCraft’s own local development stack, read on 20 October 2026: every completed website and app generation from 14 to 20 October, 13 in all. Nothing was added or removed. The caveats matter:

• n is 13, on one machine and one development database. Production runs on different infrastructure.
• Total minutes include queue time. They run from job creation to completion. The 30 minute job sat behind another job that was cancelled.
• The pipeline changed mid sample. Jobs before 17 October were a single agent run; from 17 October the multi step pipeline replaced it. The two groups measure different software.
• Stage timings were not recorded separately. The records hold the total time and each build, not the minutes spent planning, building, reviewing or fixing, so this post does not split them.

Single run versus pipeline

The four single run websites took 2.6, 2.8, 2.9 and 2.9 minutes, each with one or two builds, all green. The five pipeline websites took 4.8, 5.8, 12.4, 13.6 and 30.0 minutes, with one to five builds each. Apps took 3.3 minutes twice as single runs, and 10.5 and 16.0 minutes through the pipeline.

Separate recorded test runs point the same way. Ten one shot website generations recorded with test fixtures on 12 October took 51 to 178 seconds, about 100 seconds on average, and every one was green on its first build. Whole pipeline app generations recorded on 19 October took 439 to 504 seconds, about seven and a half to eight and a half minutes, through a plan, a foundation, two build steps, a review and a fix.

What the extra minutes buy

A pipeline run does what a person checking the site would do, in order:

1. Plan. The agent decides the pages, sections, shared components, the features to build and a design direction.
2. Foundation. It writes the home page and a guide for the rest of the site.
3. Build steps. It writes the other pages four at a time, each typechecked.
4. A green build. The site must pass a type check, ESLint, the production build and a layout check that opens every page in headless Chromium at 375, 768 and 1280 pixels and fails on sideways overflow or runtime errors.
5. Review and fix. A fresh agent reads the code and the screenshots and returns up to 12 fixes, for at most two rounds.

Each build in the sample averaged 17 to 29 seconds, except the queued job at 50. A pipeline run with five builds spends two minutes on building alone, before any thinking time. That is the price of catching a broken phone layout before you do.

Why apps take longer

An app generation runs the same stages with screens instead of pages: up to 14 screens across 3 to 5 tabs, ending with an Account tab. Its checks differ too. Instead of a website build, each round exports the app and takes screenshots of every tab at 390 and 768 pixels wide, plus the Home screen in dark mode, and the review reads those. More screens and two screen widths per tab mean more to look at, which helps explain why the two pipeline app jobs in the table took 10.5 and 16.0 minutes, against 4.8 to 13.6 for pipeline websites leaving out the queued job.

What it means for you

If you are comparing AI website builders on speed, ask what happens in the minutes you are waiting. A builder that returns a site in under a minute is usually filling a layout; one that takes ten may be writing code and testing it. Neither is better in itself: a quick draft suits trying ideas, and a checked build suits something you will publish.

Three habits keep your own wait short. Write the brief once, completely: business name, pages, features and the main action, as the AI website generator page shows. Ask for one change per chat message after generation, since each runs its own build. And use version history rather than regenerating when a change goes wrong, since restoring a known good version needs one build instead of a whole new generation. Every chat change on OneCraft is a commit, and a change whose build fails is rolled back to where the turn started, as the edit with AI chat page describes.

For how this sits among other builders, see AI website builders compared, and for what generation costs in credits, see what an AI written website costs to run.