Base44 introduces Base1, its proprietary AI model for vibe coding.

      Base44 dedicated its first year to developing applications using other companies' AI models. Now, it has created its own model, believing that ownership of the model is crucial for survival in the evolving coding landscape.

      The Tel Aviv-based startup enables users to create functional apps by describing them in simple language. It has introduced its inaugural proprietary AI model, named Base1, which it claims is already operational and serving users. Base44 asserts that it is the first app creation platform to deliver its own model instead of leasing one from a leading lab.

      Base44 trained Base1 using a dataset generated from millions of genuine user interactions on its platform. This data is significant, as every app created helps the model learn what constitutes a successful outcome. This is a feedback loop that Base44 owns, unlike its competitors.

      Base1 serves as a general-purpose agent capable of conversing and writing code. It manages multi-turn requests, utilizes tools, and oversees backend operations, similar to agent-like tools emerging across other creative platforms.

      Founder and CEO Maor Shlomo characterized the model's launch as a step toward independence. “Having our own model allows for ongoing improvements,” he remarked, reducing Base44's reliance on external providers. He was candid about the challenges involved, emphasizing that it is “not a trivial effort” requiring substantial expertise, robust infrastructure, and ample data, which many companies do not possess.

      Base44's rapid ascent is notable in the tech industry. Founded in 2024, it was sold to website giant Wix for $80 million just six months later, with a team of eight. The platform has since expanded quickly, surpassing $150 million in annualized revenue in May, following its rise past $100 million.

      Creating a model is both costly and complex, so why pursue it? The straightforward answer is financial control. Renting a leading model from OpenAI or Anthropic incurs fees for every token generated by a user's app. Shlomo stated to TechCrunch that using their own model should ultimately be “faster and cheaper for customers than relying on models like Opus.” For Base44, owning the technology grants direct control over computing and data processing costs, which is expected to yield a “structurally stronger margin profile over time.”

      This focus on margins is particularly relevant now. Parent company Wix recently downsized by 20% due to pressures from AI competition and a strong shekel. Improved margins from Base44 would be beneficial, unlike Base44, which is actively hiring.

      The launch occurs amid ongoing debates in the AI sector. If a company's operations depend entirely on another's model, how defensible is that business? A GUI based on GPT or Claude can be replicated, and the lab behind it could always shift higher up the stack. Base44 now aligns with a trend of companies developing their own models.

      That risk is not merely theoretical. Anthropic’s Claude Code has become an effective tool for app creation, providing the lab with insights into user behavior. Both Cursor and Grok’s parent company, xAI, are now integrated into SpaceX, indicating that leading labs are encroaching on Base44’s domain, armed with significantly greater resources. Platforms like Figma are also racing to implement their own AI features.

      Shlomo's strategy emphasizes specialization. He contends that frontier models “will remain very general,” while a model specifically designed for app building could outperform them in that dedicated role. He portrays Base44 as the “only vertically integrated” vibe-coding platform, owning its data, distribution, and infrastructure simultaneously. Jonathan Userovici, a general partner at VC firm Headline, describes these three elements as essential for defensibility.

      However, skepticism remains. Userovici cautioned against a weak assessment of frontier models and cited legal-AI startup Harvey, which abandoned its model training initiatives. Leading labs continue to advance, and a sufficiently competent general model could undermine a specialized one that required considerable effort to develop.

      Scale is also a fundamental factor. Base44's $150 million revenue is substantial, but Swedish competitor Lovable, which relies on external models, claims to have reached $500 million. Base1 is in its initial version, and Base44 acknowledges it is currently only “demonstrating competitive performance” against other models, instead of surpassing them. The company is betting that its advantage in data will grow faster than its rivals can catch up.

      The broader implication of this scenario is noteworthy. A year ago, relying solely on GPT was a feasible approach, but now successful AI application companies face the necessity to own their models, data, and costs. The expense of AI operations and the capabilities of major labs have made the rental model appear vulnerable. Base44 is the first vibe-coding platform to make this transition, but it likely won't be the last.