Base44 introduced its own AI model to compete with offerings from Lovable, Replit, and Cursor.

Maor Shlomo, founder and CEO of Base44, explained that the company trained its own LLM partly to move away from the generic look often associated with vibe-coded sites.

Experts have observed that websites created with leading models such as Anthropic's Opus 4.8, OpenAI's GPT-5.5, or Gemini 3 frequently share common traits including rounded corners, beige tones, and frequent emoji placement. Rivals including Lovable and Replit rely on these base models.

Paul Bakaus, CEO of Impeccable, noted in an interview that such designs resemble an algorithmic version of everyday retail stores: practical yet unremarkable.

Shlomo indicated that upcoming versions of Base 1 will aim to generate distinctly different appearances each time, though achieving this will require additional development and the team is not there yet.

He also noted that the model should deliver results more quickly while consuming fewer credits than frontier alternatives. Base44 operates as a subsidiary of Wix.

To evaluate performance, the same website was built using both Base 1 and Anthropic's Opus 4.8.

Making a website for a fictional mood lamp

The task involved creating an e-commerce page for a made-up product: an emotional support lamp that syncs with Google Calendar and shifts color based on the current event. Yellow would signal focus time, blue would indicate rest, pink would mark meetings, and red would highlight approaching deadlines.

The prompt was submitted to Base44 without any design instructions to observe the model's independent choices.

Base 1 was selected before generation. Access to this option requires a paid Builder plan.

After processing, Base44 returned the site for the product named Lumos.

Base 1 vs Opus 4.8

The initial design stood out with a deep blue background and bright yellow accents, differing from the lighter palettes common in similar projects. Standard vibe-coded elements such as rounded corners and emoji usage remained visible.

As noted by Shlomo, further refinement is needed before Base 1 consistently produces truly distinctive designs.

The identical prompt was then run again using Opus 4.8 for comparison.

Generation with Opus 4.8 took noticeably longer. The resulting site featured the expected beige background.

Beyond the color scheme, the layouts were largely alike. Base 1 completed the task more quickly, yet both consumed 1.2 message credits.

When asked to add extra color modes (green for exercise and teal for eye breaks), Opus 4.8 used 1.4 credits while Base 1 completed the update with 1.2 credits and in less time.

The credit difference was modest, but Opus 4.8 showed higher usage for the same change.

Overall, the two models performed similarly, with the main distinction being that the Opus version appeared slightly more conventional.

Shlomo said the team plans to apply reinforcement learning to encourage Base 1 to produce fresher and more varied designs.

Future versions will be reviewed upon release.