We pitted Base 44's new AI model against Anthropic's to build the same website. One was faster.

A New Challenger in AI Website Building

Base44, the no-code app-building platform that Wix acquired last year, has rolled out its own proprietary AI model for generating websites — and it's making a pointed claim: faster generation, lower credit consumption, and better visual output than relying on frontier models like those from Anthropic. A recent head-to-head test pitted Base44's new model against an Anthropic-powered build of the same website, with speed emerging as the clearest differentiator.

Why This Matters Beyond One Benchmark

The test itself is narrow — one site, one comparison — but it points to a broader shift happening across the AI coding assistant landscape. For much of the past two years, tools like Cursor, GitHub Copilot, and various app builders have functioned largely as wrappers around frontier foundation models from OpenAI, Anthropic, and Google. That approach guarantees strong baseline capability but ties a product's cost structure, speed, and even personality to whichever model it's calling under the hood.

Base44's move to train or fine-tune its own model signals a strategic pivot: rather than remaining a thin layer atop someone else's LLM, the company is betting that a purpose-built model — optimized specifically for website and app scaffolding — can outperform a general-purpose frontier model on the narrow tasks its users actually care about. That's a meaningfully different bet than what most AI coding tools are making right now.

The Credit-Burn Problem

One detail worth dwelling on is the claim about credit efficiency. Many AI-assisted coding and app-building platforms operate on consumption-based pricing, where every generation, revision, or agentic loop costs credits tied to token usage. Frontier models, especially when asked to reason extensively or make multiple tool calls, can be expensive to run at scale. A smaller, task-specific model that reaches an equivalent (or better) output with fewer tokens isn't just a speed win — it's a cost win that could reshape pricing for end users and margins for the platform itself.

What It Signals for the Wider Ecosystem

This test arrives amid growing scrutiny of AI coding tools generally, from Cursor's rapid feature iteration to a rising wave of AI code review products aimed at catching mistakes these assistants introduce. If specialized, in-house models start matching or beating frontier LLMs on narrow coding and design tasks, it could accelerate a trend where AI coding platforms diversify away from single-vendor dependency, mixing custom models with frontier ones depending on the job.

The Caveat

A single side-by-side comparison shouldn't be read as definitive proof of superiority — output quality, design taste, and reliability across edge cases matter as much as raw speed. But it's an early signal worth watching as more platforms experiment with building their own models rather than simply renting intelligence from the frontier labs.