Base44 has started rolling out Base1, a proprietary large language model trained on what the company describes as tens of millions of real user interactions from its vibe-coding platform. The Tel Aviv startup, which Wix acquired for $80 million in 2025 when it was six months old with a staff of eight, is betting that owning its own model will eventually let it beat the frontier systems it previously depended on.
The stakes are structural, not cosmetic. Vibe-coding platforms like Base44 have built their entire product on top of models from labs such as Anthropic and OpenAI, which means their margins, latency, and roadmap have always been hostage to someone else’s pricing and release schedule. Founder Maor Shlomo framed the move as a direct fix: training and owning the model “allows us a lot more optimizations on latency, cost, and efficiency.”
Shlomo does not expect Base44 to stay alone in this. He predicts other companies will train their own models too, “at least the players that have gotten enough scale and velocity to have enough data.” That is a narrower claim than it first sounds. It excludes most of the app layer and includes only firms sitting on large, proprietary interaction datasets.
Jonathan Userovici, a general partner at Headline (whose portfolio includes Mistral AI but not Base44), told TechCrunch that defensibility for AI startups rests on three ingredients: data, distribution, and tech stack. Base44’s move fits that framework directly, and Shlomo is positioning the company as what Userovici calls a fully vertically integrated player, one that owns all three at once.
The counterargument is Harvey. Userovici pointed to the legal tech startup as a cautionary case: Harvey explored training its own model and abandoned the plan, judging that frontier labs were moving too fast to out-build. Userovici does not expect most applied AI companies to become frontier labs themselves. Base44’s bet only pays off if its accumulated interaction data is a real moat, not just an expensive detour before the next frontier release does the job cheaper.
The more immediate driver may be simpler: inference cost. Userovici said enterprise customers are increasingly unwilling to pay for frontier-model performance on tasks that do not need it, which has spawned a market for orchestration tools that route requests to cheaper models without sacrificing output quality. Base44’s own press materials echo this. The company says ownership of the model gives it direct control over compute and inference spend, which it expects to translate into a stronger margin profile over time, though it did not disclose a timeline or a percentage.
Competitive pressure is also closing in from a different direction than rival vibe-coding tools like Lovable, the Swedish startup that hit unicorn status on its Series A last summer while still relying on external LLMs. Cursor and xAI now both sit under SpaceX, and Anthropic’s own Claude Code has become a vibe-coding competitor in its own right. That gives frontier labs a feedback loop into app-creation data that used to be Base44’s exclusive turf. Shlomo’s counter is that frontier models will keep improving in general capability but stay generalist by design, leaving room for a specialized, narrower model tuned specifically to app generation.
Base44 has grown headcount since the Wix acquisition even as its parent company cut 20 percent of its own workforce, and Base44 passed $150 million in annualized recurring revenue in May, up from $100 million two months earlier. That is still a third of what Lovable reported hitting this month at $500 million in ARR. Base1’s real test is not whether it launches; it is whether Base44’s growth rate holds once a specialized in-house model has to compete against whatever Anthropic ships next inside Claude Code.
For app-layer AI startups watching this, the build-versus-buy calculus now hinges on one number they should each know cold: the size and specificity of their own interaction dataset relative to the frontier labs’ general-purpose training data. Without that edge, training a proprietary model is a cost center, not a moat.
Reported by TechCrunch on June 30, 2026.
