Base44 is a tool that lets people build real, working apps by describing what they want in plain language. You type what you need, and the platform produces it. No developer required. It was bought by Wix just a year ago for $80 million, when the company had six employees and no outside funding. Since then, it has grown to over $100 million in annual revenue.
Until this week, Base44 ran on AI models licensed from OpenAI and Anthropic, the same companies behind ChatGPT and Claude. That is how almost every app-building platform works. You build a product on top of someone else's AI, and you pay them every time a user does something. Base44 has now broken from that model by launching Base1, its own in-house AI trained specifically for building apps.
This is not a model built from nothing. Building a frontier AI from scratch costs billions of dollars. Instead, Base44 took an existing open-source model and re-trained it intensively using data from its own platform: tens of millions of real user sessions, what apps people built, what worked, and what did not. The model keeps learning in simulated environments, running tasks, scoring its own outputs, and improving. Founder Maor Shlomo put it plainly: general AI models are trained to be good at everything from coding to poetry. Base44's model does one thing, building web apps, and gets better at it every day.
For now, the company says Base1 will perform on par with the best external models available for app creation. Over time, it expects the specialized model to outperform them for this specific task.
The financial logic is straightforward. Every time a user asks an AI platform to do something, the platform pays an inference cost: the price of running an AI query. That cost is one of the largest expense lines for any AI-native business. Owning the model means owning that cost. Base44's parent Wix confirmed the move is expected to produce a stronger margin profile over time, though not immediately.
The timing matters. Wix cut roughly 1,000 jobs in May, about 20% of its entire workforce, after posting a net loss of $57.5 million in the first quarter despite revenue rising 14% to $541 million. The stock has fallen over 50% in 2026. Base44 is a bright spot inside a company under serious financial stress, and any structural cost reduction at the subsidiary is relevant to the parent's recovery.
The competitive picture adds another layer. Lovable, the Swedish rival in the same space, had $500 million in annual revenue as of this month and raised $330 million at a $6.6 billion valuation in late 2025, backed by the investment arms of Alphabet and Nvidia. Lovable still relies entirely on external AI models from OpenAI and Anthropic. If Base44's cost structure improves while Lovable's stays fixed, the pricing gap between them could become a real competitive tool.
There is a genuine counter-argument. Harvey, the legal AI company often cited as a case study in domain-specific AI, abandoned plans to build its own model and moved to a multi-model strategy instead, pulling from Anthropic, Google, and OpenAI based on the task at hand. The reasoning: frontier models are improving so fast that the cost and effort of custom training may not stay ahead of what external providers offer next quarter.
Base44's answer to that is specialization. The company's founder argues that the biggest AI labs will always optimize for breadth, not depth. A model trained exclusively on app creation, with product and design intuition baked in from millions of real sessions, has an advantage that general models cannot easily replicate by simply getting smarter overall.
For business operators who use or are considering these kinds of tools, the practical signal is this: the platforms that own more of their own infrastructure will likely be cheaper and more stable to work with over time. Vendors that rent all their AI from third parties pass those costs to you, and those costs are going up. Watch for pricing differences opening between Base44 and competitors that have not made this move. That gap, if it appears, is the model working as intended.