The default model behind Ullbek has changed twice in the last year, and each time, the conversation in our team wasn't really about benchmarks. It was about what the model felt like to work with.
Most writing about models treats them like interchangeable engines. Different horsepower, different fuel economy, same shape. After months of using all the major ones in production, we think they're more like people. They have personalities. The one you want depends on what kind of work you're doing together.
Here's the cast — and what each one is actually like.
Grok 4.5 — the muscle
"Working with Grok 4.5 is like working with someone who has muscle."
When you hand it a brief, it doesn't hesitate. The work comes back fast, dense, and capable. If you ask for a sprawling landing page with twelve sections, twelve sections arrive. If you ask for something technically difficult — a complex layout, a tricky data structure, a long page that needs to keep its energy to the bottom — it just does it. The first version is rarely the prettiest, but it's rarely the wimpiest either. You feel like you're working with someone who can lift anything you put in front of them.
Where Grok 4.5 isn't the right fit is when taste matters more than volume. It's the colleague who'd rather show you three layouts than pick one and commit. Capable, present, dependable — but not the model you'd hand a brief where the answer is supposed to be a thing, not a lot of things.
Volume. Long pages, dense builds, technically heavy work where the first draft needs to come back with real weight behind it.
Sonnet — the deep mind
"Working with Sonnet is like working with a philosopher or an artist."
It's the model that reads your brief and asks you a question back. Not literally — but you can feel it. Where other models hear the words you wrote, Sonnet hears the brief underneath the brief: what you meant, what kind of site this should be, what feeling you were going for. The first version is rarely complete, but it's the version that has the most to say. Sometimes the second pass is just a polish, and sometimes the second pass is a completely different site — and the different one is the right one.
Sonnet shines on work that benefits from thought. A personal essay site. A studio's portfolio. A small brand that needs a voice, not just a layout. Where it loses time is on the volume work — the cases where you just need twelve good-looking sections that don't fight you. Sonnet is too busy thinking to be that efficient, and that's the trade.
Briefs where the answer needs a point of view. Personal sites, editorial work, anything where the model has to figure out what kind of thing it is before it can build it.
MiniMax M3 — the easygoer
"Working with MiniMax M3 is like working with the guy who takes things easy — but still lands on the money."
Not lazy — easy. They take the brief, they read it, they get to work, and the work comes back on the money. There's no overthinking. There's no showing off. The first version is usually close to right, and the second pass is usually small. You finish the build in fewer rounds than you expect to, and the work itself has a restraint to it — nothing extra, nothing missing, just the right amount of everything.
M3 is the default on Ullbek. The reason isn't that it wins every benchmark. It's that it has the highest good output per round of feedback, and that's the metric that matters when a real person is building a site. It's the model that gets out of your way the most — which, when you're trying to ship a bakery's homepage by Friday, is the highest compliment a model can earn.
Everyday site-building where you want the first version to be close, and the second version to be small. The default for a reason.
GLM — the technical mind
"Working with GLM is like working with a technical mind that uses mathematical approaches to figure things out."
GLM doesn't see images — but it doesn't need to. It reads the brief, parses the structure, and produces output that's structurally tight. Spacing that aligns. Margins that make sense. Code that's clean. When something is broken, GLM is the model that fixes it without being asked three times. It treats layout like a system to be solved, and it solves it.
GLM was our default for a while, and the reason it lost the seat to M3 wasn't because it was worse. It was because M3 also has the technical mind — plus vision, plus the ease we just described. GLM is still a model we reach for, specifically when the work is dense, structural, or coding-heavy, and the brief doesn't need visual judgment to land.
Structural work. Debugging. The parts of a build where you want the model to think like an engineer instead of a designer.
Even though M3 is the default, every model in this post is available on Ullbek — you can switch between them in your build settings. Pick the one that fits the work, the deadline, or the way you want to work today. The default is just the seat we think most people will want. The right model for you is the one that matches your taste, not ours.
Why personality matters more than the leaderboard
Benchmarks tell you what a model can do. Personality tells you what it's like to do it with.
When you build a site on Ullbek, you don't fire off one prompt and walk away. You have a conversation. You push, you redirect, you ask for a different version, you ask for a smaller change. You're working with the model for as long as the build lasts, and the shape of that conversation is the build.
A model with muscle gets you a long way — but you have to manage the volume. A model with depth gets you somewhere surprising — but you have to wait for it. A model that's easy to work with gets you a finished site in fewer rounds. A model that's technical gets you a site that doesn't break. None of these is the right answer on its own. The right answer depends on the site, the deadline, the kind of work — and increasingly, on the kind of collaborator you want on the other side of the chat.
The cast keeps changing
The default will keep moving. New models arrive, we re-evaluate, and the one we think fits the work best takes the seat — but the seat is just a default. Every model in this post is available on Ullbek, and you can switch to any of them for the build you're doing. The right model is a personality match, not a benchmark match — and the right personality might be different from ours.