Replit switches default AI provider from Anthropic to OpenAI

Replit, the AI-assisted coding platform, announced that OpenAI will replace Anthropic as its default model provider. Chief executive and co-founder Amjad Masad made the disclosure onstage at TechCrunch Disrupt, citing cost, partnership quality and performance as the drivers. Anthropic’s models remain available for users who select them, though Masad acknowledged most customers stay with the default.

Until the August agreement with OpenAI, Replit ran on Anthropic’s Claude family and ranked among the startup’s five largest customers, a relationship Masad said fueled substantial revenue growth. OpenAI is now the preferred partner for Replit Agent, the company’s autonomous coding agent, on the strength of performance, economics and partnership depth.

Masad said ChatGPT has pulled ahead and pointed to Replit’s internal benchmark as well as a public one showing OpenAI’s Sol and Astra models outperforming rivals at building applications from scratch using a “vibe-coded” approach. The financial logic is straightforward: because the majority of users never change the default, the switch stands to cut Replit’s Anthropic spend significantly.

Public artifacts tell a more recent story. A late-September blog post still listed Claude Fable 5 and 5.1 as models running in Replit Agent’s production environment, and a product team member posted on X about using Anthropic’s flagship Opus 5.5. After the conference Masad emphasized that Replit is model-neutral and that customers can choose Claude or any other supported model. Sources close to the Anthropic partnership, however, disputed Masad’s on-stage characterization and said the commercial relationship remains in force.

Masad also shared growth metrics. Enterprise deals have risen from roughly 2 % of revenue last year to about 15 % this year, and more than half of revenue now comes from B2B contracts. Privately he told attendees he expects enterprise and government customers to account for over half of Replit’s revenue by 2027. He declined to disclose precise ARR figures.

On the topic of AI guardrails, Masad drew a parallel between human workplace needs — comfortable offices, food, climate control, hardware — and what AI agents require: a “habitat” with reliable scheduling, error recovery, safety and security, isolated environments, networking, identity and permissions. Only once that infrastructure is in place, he argued, can AI be deployed safely.