Replit drops Anthropic for OpenAI, says its CEO: model loyalty is now a matter of cost and performance

By TechCrunch · October 16, 2026.

Amjad Masad, co-founder and CEO of Replit, said Thursday at TechCrunch Disrupt that his company no longer works with Anthropic. Replit is a cloud platform where applications are developed and deployed from the browser. According to Masad, the decision was driven by several factors, including cost.

"I think ChatGPT is better now," he said. He added that Replit uses an internal benchmark and publishes a public one, and that on them "Sol and Astra" perform better at building "vibe-coded" applications (AI-generated from natural-language instructions) from scratch. He concluded that it was a decision based on cost, partnership and performance, and that OpenAI is "a better partner for now."

There is an important caveat. Replit announced a partnership with OpenAI in August, but had never indicated it had broken with Anthropic. According to TechCrunch, a person familiar with the relationship, who asked not to be named, maintains that the partnership remains in place. A September blog post cites Claude Fable 5 and Fable 5.1 as models in Replit Agent's production environment. TechCrunch asked Replit for clarification and will update the story.

Masad also gave business figures, without a specific ARR (annual recurring revenue). The enterprise segment went from around 2% of revenue last year to about 15% this year, and more than 50% of revenue is B2B, in his words. This puts the pressure on cost and reliability in context: an enterprise customer demands predictability.

Our take. First, caution: for now we have a statement made on stage, a conflict with an anonymous source and a benchmark that is Replit's own. It is partisan evidence, not independent measurement. "Sol and Astra" winning on its test does not amount to overall leadership; the specific task of generating apps from scratch may favor a different model from the one that leads on other tasks.

Second, the underlying thesis: the base model is becoming an interchangeable layer. When a platform can switch providers based on price and performance on a specific task, the advantage shifts to the "habitat" Masad himself described: orchestration, error recovery, sandboxing (isolated environments), identity and authorization. Whoever controls that plumbing keeps the customer; the model lab, not necessarily.

Third, for Anthropic it is an uncomfortable but limited signal. Losing or scaling back a high-profile partner in programming, its stronghold, matters more for reputation than in figures we do not know. For OpenAI it is a useful trophy when pitching to enterprises.

In the short term, this fuels price wars and "provider switch" announcements with fine print. In the long term, it is good news: competition makes the ability to build software cheaper and brings that power to more people, one more piece on the path toward abundance. It is worth waiting for Replit's clarification before treating the break as complete.