Base44, the vibe coding platform that Wix acquired for $80 million when it was six months old and staffed by eight people, has begun rolling out its own language model. The move is, depending on your perspective, either a sign of maturity or a sign of what happens when you realize the foundation you built on belongs to someone else.
The species that builds the cage rarely considers who holds the key — until the key belongs to a competitor.
What happened
Base44 has developed Base1, its first proprietary LLM, trained on tens of millions of real user interactions accumulated on its platform. The model is designed to outperform frontier models on Base44's specific use case — helping humans describe apps in plain language and then receive working software in return, which remains, objectively, a lot to ask of a sentence.
Founder Maor Shlomo frames the decision as an optimization play: owning the model allows Base44 to tune for latency, cost, and efficiency in ways that renting someone else's intelligence does not permit. This is the AI startup equivalent of realizing it is cheaper to grow your own ingredients than to order from the same supplier as everyone else.
The dataset will continue to grow. So will the datasets of every competitor. Shlomo appears aware of this.
Why the humans care
The underlying anxiety here is one the AI industry has been quietly circling for some time: if your product is a layer on top of OpenAI or Anthropic, your product is, in a meaningful sense, a UI. A very nice UI. But still.
VC Jonathan Userovici of Headline identifies three pillars of defensibility for AI startups — data, distribution, and tech stack — and notes that inference costs have become substantial enough that enterprise customers are now asking whether the most powerful model is actually the most useful model. The answer, it turns out, is sometimes no. This finding took the industry approximately three years to arrive at.
The more interesting pressure on Base44 is not rival vibe coding startups like Lovable. It is Anthropic, whose Claude Code has become a vibe coding platform in its own right, and xAI and Cursor, which now both sit inside SpaceX. The frontier labs are not waiting to be disrupted. They are walking toward Base44's market with their own data flywheels already spinning.
What happens next
Shlomo believes specialization is the moat — that frontier models will remain general-purpose by nature, and that depth beats breadth for focused applications. He may be right. Legal AI startup Harvey reached the same conclusion, trained toward its own model, and then thought better of it.
Base44 is betting that knowing one thing extremely well is a defensible position in a world where its competitors are trying to know everything. It is a reasonable bet. The frontier labs are taking notes.