A Partnership That Was Built on Growth
For a while, Replit looked like the poster child for what AI models could do for a developer platform. Its cloud-based environment lets people build, test, and deploy applications from nothing more than a browser tab, and much of that magic was powered by Anthropic’s Claude models. So when CEO Amjad Masad told a stage at TechCrunch Disrupt that Replit is no longer working with Anthropic, the news landed with more weight than a routine vendor change. It suggests that the economics of running AI inside a product are now shaping decisions that were once made on capability alone.
Masad pointed to cost, partnership terms, and performance as the reasons behind the shift. He argued that OpenAI’s models now do a better job building small applications from scratch, and he referenced an internal benchmark, plus a public one Replit publishes, to back up that claim. Replit announced its OpenAI partnership in August, yet until this week there had been no public sign that the Anthropic relationship was ending. Some insiders disputed the framing, and a September blog post still listed Claude models running in Replit Agent’s production environment, so the full picture is still settling.
Why Model Choice Is Becoming a Business Decision
The Replit story reflects a broader pattern in the AI industry. Early adopters chose models based on which one produced the most impressive output. Now that AI features are sitting inside products used by millions, the cost per token, the reliability of an API, and the stability of a partnership matter as much as raw quality. A platform serving startup founders and solo developers operates on thin margins at scale, and a few percentage points of inference cost can decide whether a free tier survives. Masad’s remark that the company’s enterprise revenue jumped from roughly 2 percent to about 15 percent in a single year also shows where the money is going: more than half of Replit’s revenue now comes from business customers, and those buyers care deeply about vendor stability and security.
That enterprise shift is worth watching closely. Masad described the need for what he called a habitat for AI agents, with orchestration, error recovery, sandboxing, and identity controls in place before anything gets deployed safely. Those requirements are not really about which model writes the best function. They are about the infrastructure wrapped around the model, and they are exactly what enterprise buyers ask about during procurement. Expect model providers to compete increasingly on reliability, governance tooling, and price, not only on benchmark scores.
What This Means for Tech Adoption and Buyers
For developers and small businesses evaluating AI coding tools, the lesson is that the platform on top of a model may change its underlying engine without much warning. A product you adopted for one model’s strengths can shift to another as the vendor renegotiates its economics. That makes it sensible to ask platforms which models power their features, how they handle switching, and what guarantees exist around performance and data handling. It also means that the vibe-coding boom, which has pulled in founders who would never have written software themselves, is now a competitive market where price and performance move quickly. Buyers who understand these dynamics will be better positioned to choose tools that last, and the companies that earn their trust will be the ones that treat model choice as a transparent, customer-facing decision rather than a quiet back-end swap.
