Replit CEO says company no longer works with Anthropic, favors OpenAI

Replit, the browser-based AI coding platform popular with startup founders and developers, is no longer working with Anthropic, according to its CEO and co-founder Amjad Masad. He made the remark onstage at TechCrunch Disrupt on October 15, 2026.

Masad said several factors drove the decision, cost among them. Replit unveiled a partnership with OpenAI in August, but gave no sign at the time that it had stopped using Anthropic. Masad told the audience he believes ChatGPT is now the better option.

The company relies on an internal benchmark to track metrics such as performance and also publishes a public one, Masad said. He claimed that Sol and Astra consistently outperform on building vibe-coded applications from scratch. "It was a decision based on both cost, partnership, and performance," he said.

Conflicting accounts

Not everyone agrees with Masad's characterization. A person familiar with the relationship between the two companies, speaking on condition of anonymity, said the partnership was still active. A Replit blog post from September also listed Anthropic's Claude Fable 5 and Claude Fable 5.1 as models running in the production environment of Replit Agent. Replit has not yet clarified the discrepancy.

Under its arrangement with Anthropic, Replit had used Claude models to power its platform, a tie-up credited with helping the company grow revenue substantially. A genuine split would mark a notable shift in the competition among model providers for the fast-growing AI coding market.

Enterprise share climbing

Masad declined to disclose annual recurring revenue but offered a breakdown of growth. He said enterprise contracts made up roughly 2% of revenue last year and around 15% this year, while more than half of Replit's revenue now comes from business customers.

He also addressed AI safety, arguing that agents, like human workers, need a suitable environment to do good work. That "habitat," he said, requires strong orchestration, error recovery, security and sandboxing, networking, and identity and authorization controls before AI can be deployed safely.