We Tried Base44’s First LLM Model to Make a Website

Base44 launched its own AI model to edge out competitors like Lovable, Replit, and Cursor.

Maor Shlomo, the founder and CEO of Base44, told Business Insider that part of the reason the San Francisco-headquartered vibe-coding startup trained its own LLM was to reduce the AI-slop look that has come to define vibe-coded products.

Design gurus have said that websites built with frontier models like Anthropic’s Opus 4.8, OpenAI’s GPT-5.5, or Gemini 3 often share similar interfaces and design elements, such as rounded corners, beige color palettes, and excessive use of emojis. Base44’s competitors, such as Lovable and Replit, use these underlying models.

For instance, Paul Bakaus, the CEO of AI design startup Impeccable, said in a June interview with Andreessen Horowitz that the designs of these sites look like an “algorithmic Uniqlo or Ikea” — functional but basic.

Shlomo said Base 1, in its more advanced versions in the future, will “create something that looked uniquely different” every time, but “It’s going to take some effort, so we are not yet there.”

He also said the model would produce faster results and cost users fewer credits than using frontier models. Base44 is a subsidiary of the website-building company Wix.

To see how the model fared, I pitted Base 1 against Anthropic’s Opus 4.8 to vibe code the same website.

Making a website for a fictional mood lamp

I settled on building an e-commerce website for a fictional product — an emotional support lamp that can read Google calendars and change its hue depending on the ongoing event. I asked it to glow yellow for focus mode, blue for downtime, pink for meetings, and red for impending deadlines.

I fed Base44 my prompt, carefully avoiding any instructions about website design, curious to see what design it would come up with on its own.

Before pressing Enter, I selected Base 1 as my AI model. Free users don’t have this option; only users with the $40-per-month “Builder” subscription or higher can toggle between AI models.

After a few minutes of anticipation, Base44 presented me with the site for “Lumos.”

Base 1 vs Opus 4.8

Off the bat, I was surprised by the design. The website had a deep blue background, which already set it apart from the light or beige-colored front ends I’d seen on most vibe-coded websites. It was complemented by a bright yellow accent color.

The rest of the website was basic, with some telltale signs of being vibe-coded. For instance, the elements had rounded corners, and there were plenty of emojis used.

As Shlomo said, Base 1 has a long way to go before its designs can be truly unique.

For comparison’s sake, I fed Base44 the same prompt again, but chose Anthropic’s Opus 4.8 as the working model this time.

The first thing that struck me was that Opus 4.8 took longer to generate the site than Base 1, which felt like twice as long. When it finally loaded, I had my first “aha!” moment — hello, beige background.

Apart from the color palette, everything else was similar to the Base 1-coded website. While Base 1 was faster, both used the same number of credits, 1.2 message credits each out of my 250 monthly credits.

Then I tried tweaking both websites to add one extra feature. I prompted both models to add a few more color modes — green for exercise and teal for eye breaks. To complete this task, Opus 4.8 burned through 1.4 message credits, while Base 1 got the job done in 1.2 credits and made the change faster.

While the difference was small, Opus was proving to use credits faster for what was effectively the same tweak.

All things considered, I did not see a major difference in how the two models worked, aside from the Opus-generated design looking a tad more generic.

Shlomo said the Base44 team will conduct “reinforcement learning” on Base 1, which involves prompting the model to keep generating designs that look new and unique.

We’ll check back in when the next version is released.