Replit Shifts Default Model to OpenAI Amid Enterprise Push

Replit is realigning its model dependencies, pivoting to OpenAI as the default provider for its AI coding platform, as TechCrunch first reported. During a presentation at TechCrunch Disrupt, Replit CEO and co-founder Amjad Masad announced that the company is no longer working with Anthropic, citing a combination of cost, partnership strength, and performance metrics.

According to TechCrunch, the shift is particularly significant for the Replit Agent, which will now be powered by OpenAI. Masad noted that internal and public benchmarks indicated that OpenAI's Sol and Astra models performed better when building "vibe-coded applications from scratch." While Masad previously indicated the company was no longer working with Anthropic—a firm Replit was once a top-five customer for—he later clarified to TechCrunch that the platform remains model agnostic. Users can still manually select Anthropic's Claude models, though Masad noted that most customers typically stick with the default setting.

This strategic pivot coincides with a broader push toward enterprise scalability. Masad told TechCrunch that B2B revenue now accounts for more than 50% of the company's total revenue. Enterprise contracts specifically grew from approximately 2% of revenue last year to 15% this year. Looking ahead, Masad expects enterprise and government customers to drive more than half of Replit's revenue by 2027.

However, the nature of the break with Anthropic remains a point of contention. TechCrunch reports that insiders familiar with the partnership rejected Masad's on-stage comments, claiming the partnership is still in place. This is supported by Replit's own marketing materials; a late-September blog post cited Claude Fable 5 and 5.1 as active in the Replit Agent production environment, and a product employee recently promoted the capabilities of the Opus 5.5 model on X.

Beyond model selection, Masad emphasized the necessity of a "habitat" for AI agents to ensure safe deployment. He told TechCrunch that this environment must include sandboxing, identity and authorization, networking, and robust error recovery to allow agents to function effectively.