Why we made GLM 5.2 the default model

Choosing the model that powers an AI website builder is not a benchmark exercise. It's not about which model scores highest on a coding leaderboard or which one is cheapest per million tokens. It's about a single, stubborn question: when a non-technical person types one sentence, does the model produce a website a designer would respect?

This post is the record of one such evaluation — the exact prompt, the site our current default model produced from it, and what the output tells us. The default will keep changing as new models arrive and existing ones improve; that's a moving target, and we'll keep re-evaluating. But the method stays the same: give the model a real prompt, show what comes out, and judge it openly.

The prompt — one sentence, no follow-ups

Here is the entire prompt. Nothing was added, clarified, or iterated. No reference images, no brand kit, no second message. This is what went in:

// Single prompt, no follow-ups, no reference images:
Create a striking, minimalist portfolio website for 'Elias Voss Photography' specializing in fine art landscapes and portraits. Monochrome + selective color aesthetic (deep blacks, crisp whites, and muted teal accents). Large, dramatic full-screen images with perfect typography overlay. Smooth gallery navigation, about page with artistic bio, print shop with elegant framing options, and journal. The site must feel premium, emotional, and gallery-like.

That's it. One paragraph. Six sentences describing a fictional fine art photographer — the aesthetic, the pages, the feeling. No wireframe, no colour hex codes, no font names, no layout instructions. The model had to infer all of it.

This is the test that matters, because it's the test that mirrors reality. Most people who use Ullbek don't write detailed briefs. They write something like this — a feeling, a vibe, a few nouns. The model has to do the rest.

The site it built — embedded live, right here

This is not a screenshot. This is the actual published site, loaded live in the frame below. Click around it — every page, every link, every gallery transition is real. It was built and published entirely by GLM 5.2 from the single prompt above.

If you're on a phone, the frame above is compact — tap Open ↗ to view the full site in a new tab. Either way, what you're seeing is the unedited output. No human touched the HTML, the CSS, the images, or the copy after the model finished.

What the model actually built — the structure

From that one paragraph, GLM 5.2 made a six-page site with a coherent information architecture, a consistent visual system, and original copy written in the voice of a real fine art photographer. Here's what it produced:

Notice what's not there: no placeholder text, no "Lorem ipsum," no broken links, no empty pages. The model invented a coherent artist — Elias Voss, working from the north coast since 2009 — and wrote every word of copy in that voice. The exhibitions are fictional but plausible. The print editions are numbered and priced consistently. The journal entries read like a real photographer's field notes.

Two sides of the same coin — the harness and the model

It's worth being precise about what this test actually proves, because it proves two things at once.

The model — in this case, GLM 5.2 — proved it can take a vague, emotional brief and produce a website with aesthetic judgement, not just code. That's the side most people focus on when they talk about AI website builders: is the model smart enough?

But the model alone can't build a website. It can write code. It can't see the result. It can't check whether the hero image actually fills the screen, whether the gallery layout breaks on a phone, whether the CSS has a broken rule, whether a link 404s. That's the harness — the agent layer that wraps the model. It gives the model the tools to write files, place images, verify the render, fix its own errors, and publish.

This is why we don't talk about the model in isolation. When you build a site on Ullbek, you're not just talking to an LLM. You're talking to a system — a model with taste, wrapped in an agent that can see and fix and ship. The Elias Voss site is proof that the combination works. The model wrote the code; the harness made sure it actually rendered, published, and stayed live.

Why we're showing our work

Most companies pick a model and don't tell you why. They might say "powered by AI" or name a model in a footnote. They won't show you the prompt or the output.

We're doing the opposite. The prompt is above, in full, in a code block you can copy. The site is embedded above, live, in a frame you can click around. The evaluation is visible.

Because this is the question that actually matters to anyone using an AI builder: not "is it AI?" but "is it good?" And the only honest answer is to show you.

So here's the invitation. Take the prompt above — copy it verbatim. Paste it into Ullbek and see what our current default model builds for you. The default will change over time — new models arrive, existing ones improve — but the method won't: real prompt, real output, shown openly.