Base44, the platform for vibe coding that was purchased by Wix for $80 million just a year ago — at a time when the venture was only six months old with a mere team of eight — has begun the rollout of its proprietary AI model designed to assist users in developing applications using natural language.

This initiative arises amid heightened discussions in AI communities regarding the suitability of frontier models for various applications. A pertinent inquiry is whether businesses utilizing external models can maintain long-term defensibility. Base44’s latest action, headquartered in Tel Aviv, addresses both of these concerns.

Although its custom LLM is in the early stages of deployment, Base44 aspires for it to eventually outshine frontier models. The founder, Maor Shlomo, stated, “training and owning the model as part of [our] entire stack grants us considerably more optimizations regarding latency, cost, and efficiency.”

On the surface, this may be a strategy to maintain an edge over competitors such as the Swedish startup Lovable, which achieved unicorn status during its Series A funding last summer and relies on third-party LLMs. Nevertheless, Shlomo anticipates that others will train their own models — “at least those entities that have scaled sufficiently and generated enough velocity to gather ample data.” 

Jonathan Userovici, a general partner at VC firm Headline — whose portfolio encompasses AI companies like Mistral AI, but not Base44 — asserts that data is one of three essential components of defensibility for AI startups, alongside their distribution and technology stack. 

The result is that firms with robust brands are now leaning into their data and frameworks to bolster their defensibility, with Base44 exemplifying this trend. The company claims that the initial iteration of its LLM, Base1, was created and trained using a dataset derived from “tens of millions of genuine user interactions on the platform.” 

This dataset will continue to grow alongside the company; however, so too will that of its competitors. The most significant rivalry may not emerge from vibe-coding startups, but from frontier AI laboratories encroaching on Base44’s territory — both Cursor and Grok’s parent company, xAI, are now part of SpaceX, and Claude Code has emerged as a vibe coding contender in its own right. 

This allows Anthropic and other foundational AI providers to access data and feedback mechanisms to enhance models for application creation, yet Shlomo contends that specialization provides Base44 with a competitive advantage. “Models are evolving, but they’ll remain quite general in their capabilities,” he predicted.

Userovici urges caution against underestimating frontier models, referencing the legal tech startup Harvey, which reversed its decision to develop its own model. He does not foresee applied AI companies collectively transitioning into frontier labs but contextualizes Base44’s initiative within a wider framework — where inference costs have become a pivotal part of the landscape.

Userovici notes that this cost pressure has instigated changes now being demanded by enterprise customers. “They don’t necessarily perceive a [return on investment] when utilizing the latest models for all use cases, leading to the establishment of entire infrastructures aimed at orchestration and optimization to select the appropriate models for them, ensuring costs don’t escalate while preserving similar performance across the majority of applications.”

Although enterprise companies still constitute a minority among users of vibe coding platforms, they represent a growing portion of revenue, and users of all sizes are beginning to voice concerns over AI usage costs. Base44’s decision to craft its own LLM is influenced by various factors, with cost reduction likely being among the advantages.

“We aim to develop a model that’s more aligned with our vision, optimized to reflect what users appreciate in terms of the outcomes we’re achieving, and is ultimately quicker and less expensive for customers compared to using frontier models like Opus,” Shlomo remarked.

Concerning Base44, the path toward cost reduction isn’t straightforward. In a press release, the company clarified that “ownership of the model grants Base44 direct authority over compute and inference expenses, which is anticipated to yield a structurally stronger margin profile over time.” 

Even with a delayed benefit, enhanced margins would be advantageous for Base44’s parent company, which has recently announced layoffs affecting 20% of its workforce. Conversely, Base44 has been increasing its headcount since the acquisition — and disclosed that it surpassed $100 million in annual recurring revenue a few months ago.

That still trails Lovable, which announced hitting $500 million in ARR earlier this month. However, Shlomo is confident that the “significant engineering effort” involved in developing Base1 will solidify Base44’s status as the “only vertically integrated vibe-coding application — meaning, in Userovici’s view, a player that possesses its distribution, data, and infrastructure simultaneously.

This article was updated to clarify Base44’s location.