Replit, the browser-based development platform that offers AI-assisted coding, announced on the main stage of TechCrunch Disrupt that it is no longer working with Anthropic. The statement came from CEO and co-founder Amjad Masad, who confirmed the termination of the collaboration that had previously allowed Replit to run Anthropic’s Claude models within its service. The partnership had been credited with contributing to significant revenue growth for the company.
Masad explained that the split was driven by several factors, chiefly cost, performance and the overall quality of the partnership. He said internal benchmarks showed ChatGPT outperforming Anthropic’s offerings and that a public benchmark consistently ranked Replit’s Sol and Astra models higher for building vibe-coded applications from scratch. “I think ChatGPT is better now,” he told the audience, adding that the decision reflected both financial and technical considerations.
In August, Replit announced a new partnership with OpenAI, but Masad made no indication that the Anthropic relationship had been terminated at that time. He framed the shift as moving to a partner that better aligns with the company’s current needs, emphasizing that the change was based on a combination of cost efficiency, performance metrics and strategic fit. The announcement did not include any details about the status of the Anthropic collaboration.
Despite Masad’s statements, a source familiar with the arrangement, who requested anonymity, claimed the Anthropic partnership remains active. The company’s website, recent blog entries, and social-media posts continue to reference Anthropic models. A late-September blog post listed Claude Fable 5 and Claude Fable 5.1 as running in Replit Agent’s production environment, while a product team member posted on X about using the flagship Opus 5.5 model and promoting its capabilities.
Masad also provided an update on Replit’s revenue composition, noting that enterprise contracts have grown from roughly two percent of total revenue last year to about fifteen percent this year. He added that more than half of the company’s income now comes from business-to-business customers, although he declined to disclose annual recurring revenue figures. The shift reflects a broader strategy to target larger organizational clients.
He concluded with a description of the environment required for AI agents to operate safely. According to Masad, agents need a “habitat” that provides reliable orchestration, error recovery, security sandboxing, networking, identity management and authorization. He argued that once these components are in place, developers can deploy artificial-intelligence services with confidence in their safety and reliability.