Base44 builds own AI model to escape frontier dependency

Key Takeaways

• Base44 launched Base1, a custom LLM trained on tens of millions of user interactions from its vibe-coding platform
• The move reflects growing concern that AI startups built on third-party models lack long-term defensibility
• Base44 hit $150M ARR in May 2026, two months after crossing $100M, though rival Lovable claims $500M ARR

Base44, the Tel Aviv-based vibe-coding platform Wix acquired for $80 million last year, is rolling out its own AI model. The company trained Base1 on tens of millions of user interactions, a bet that owning the model will deliver better latency, lower costs, and a moat competitors cannot easily copy.

The launch arrives as a growing number of AI startups question whether building on Claude, GPT-4, or other frontier models is sustainable. If Anthropic or OpenAI can enter your vertical tomorrow with superior data and distribution, what exactly do you own?

Why Base44 built its own model

Founder Maor Shlomo frames the decision in terms of control. "Training and owning the model as part of [our] entire stack allows us a lot more optimizations on latency, cost, and efficiency," he told TechCrunch. Base44's platform lets users describe apps in plain language and generates working code, a workflow that demands tight integration between the model and the product experience.

The first version of Base1 draws on data generated from real user sessions. That dataset grows as more people build apps on the platform, a flywheel that could widen the gap with competitors over time.

Shlomo expects rivals to follow the same path. "At least the players that have gotten enough scale and velocity to have enough data" will train their own models, he predicted. Swedish unicorn Lovable, which hit $500 million in ARR earlier this month, still relies on external LLMs. How long that lasts is an open question.

The defensibility debate in AI startups

Jonathan Userovici, a general partner at Headline (which backs Mistral AI but not Base44), identifies three pillars of defensibility for AI companies: distribution, data, and tech stack. Owning a model addresses the third, while the interaction data addresses the second. Distribution, for now, comes from Wix's reach and Base44's own ARR momentum.

Userovici warns against underestimating frontier labs, citing legal tech startup Harvey, which scrapped plans to train its own model after deciding the effort was not worth the payoff. He does not expect most applied AI companies to become frontier labs. But he sees Base44's move as a response to a real economic shift: inference costs now matter enough that enterprise customers are pushing back.

“They don't necessarily see a [return on investment] when using the latest models for all use cases, so an entire infrastructure is being set up to do orchestration and optimization to select the right models for them so that costs don't skyrocket while maintaining the same or similar performance across the majority of use cases.”

— Jonathan Userovici, General Partner at Headline

Frontier labs are eyeing vibe coding

Base44's real threat may not be Lovable or other startups. Cursor and Grok's parent company xAI now both belong to SpaceX. Claude Code has become a vibe-coding tool in its own right, giving Anthropic access to the same feedback loops Base44 relies on.

Shlomo argues specialization is the defense. "Models are progressing, but they'll stay very general in what they can do," he said. A frontier model optimized for everything may underperform a model fine-tuned for one specific workflow. Whether that holds as frontier capabilities improve is the central gamble.

Revenue growth and margin pressure

Base44 announced $150 million in ARR in May, just two months after crossing $100 million. The company has grown headcount since the acquisition, even as parent company Wix laid off 20% of its workforce.

Owning the model could improve margins over time. Base44 says ownership gives it "direct control over compute and inference spend, expected to result in a structurally stronger margin profile." The payoff is not immediate; training and maintaining a model costs money. But if Base1 eventually replaces calls to Claude Opus or GPT-4, the savings compound.

Shlomo calls it a "huge engineering effort" and frames the result as making Base44 the "only vertically integrated vibe-coding application." In Userovici's terms, that means owning distribution, data, and infrastructure at once.

What this means for the vibe-coding market

The vibe-coding space is consolidating around a few well-funded players: Base44, Lovable, and the frontier labs themselves. Enterprise customers, still a minority of the audience, are growing as a share of revenue. They care about cost predictability and data security, two areas where owning the model helps.

Startups considering this path should weigh the trade-offs. Training a model demands data, capital, and ML talent. Relying on external models keeps you flexible but exposed. The right answer depends on your scale. Shlomo's advice: you need "enough data" before the switch makes sense.

Frequently Asked Questions

What is Base44's new AI model?

Base1 is a proprietary LLM trained on tens of millions of user interactions from Base44's vibe-coding platform. It is designed to generate app code from natural language prompts with better latency and lower cost than frontier models.

Why did Base44 build its own model instead of using GPT or Claude?

Base44 wants more control over optimization, cost, and the user experience. Relying on third-party models creates dependency and limits margin improvement. Owning the model also builds a defensible moat.

How does Base44 compare to Lovable?

Base44 hit $150M ARR in May 2026, while Lovable claims $500M ARR as of June 2026. Lovable still uses external LLMs; Base44 now has a proprietary model. Both target the vibe-coding market.

Is Base44 owned by Wix?

Yes. Wix acquired Base44 for $80 million in 2025, when the company was just six months old with eight employees.

Will other vibe-coding startups build their own models?

Base44's founder expects competitors with enough scale and data to follow. The timing depends on each company's resources and strategic priorities.

Source: TechCrunch / Anna Heim