Replit makes OpenAI its default instead of Anthropic: the model is interchangeable, the value is in the agent

By TechCrunch · October 17, 2026.

Amjad Masad, CEO and co-founder of Replit, said on stage at TechCrunch Disrupt that the platform no longer works with Anthropic. Days later he qualified the message: customers can still choose Claude models, but OpenAI becomes the default and “preferred” provider for Replit Agent. Replit is a cloud platform where applications are built and deployed from the browser.

Masad justified the decision on cost, partnership and performance. “I think ChatGPT is better now,” he said. According to him, Replit's internal benchmark and a public one show that the Sol and Astra models perform better when building “vibe-coded” applications from scratch, that is, generated from natural-language instructions. These are the company's own figures, not an independent measurement. He also said Replit was until recently one of Anthropic's five largest customers.

The story has cracks. According to TechCrunch, a source close to the relationship between the two companies, who asked for anonymity, maintains that the partnership is still in place. In addition, Replit cited Anthropic models (Fable 5 and Fable 5.1) in a late-September blog post as part of its production environment, and a product employee was promoting Opus 5.5 on X around the same time. Replit had announced its deal with OpenAI in August, without indicating then that it was ending its collaboration with Anthropic.

The detail that weighs most economically: most users do not change the default model. Moving the default could substantially cut Replit's spending with Anthropic, even if Claude remains available in the menu.

Masad also shared business figures, without disclosing annual recurring revenue (ARR). Enterprises went from 2% of revenue last year to about 15% this year, and more than 50% of revenue is already B2B. He expects more than half to come from enterprise and government customers in 2027.

**Our take.** This story is less a breakup than a symptom. Replit calls itself “model-agnostic”, and that phrase matters more than who wins this quarter. When frontier models perform similarly and prices fall, the default provider is a negotiating decision: it is renewed based on cost and performance measured on one's own task. What protects Replit is not Claude or GPT, but its agent layer, its usage data and its relationship with the customer.

For the labs, the lesson is uncomfortable. A large, loyal customer can shift its traffic with a configuration change, and bargaining power moves toward whoever controls the end experience. Anthropic and OpenAI are also competing to become preferred partners of these platforms, not just API providers.

The messaging should be treated with caution. The benchmark backing the switch is Replit's own, the CEO has commercial incentives to present it as a quality decision, and accounts of the real state of the relationship contradict each other. Until there are independent measurements, “X is better now” is a self-interested claim.

Masad also put forward an idea worth noting: agents need a “habitat” with good orchestration, error recovery, sandboxing, identity and authorization. It matches our underlying thesis: the real risk and control of agentic AI are decided in that infrastructure, not in the model. In general, the more interchangeable the model, the more important it will be who governs that environment.

In the short term this means more commercial volatility and price pressure among labs. In the long term it is good news: competition lowers the cost of access to tools that let more people create software without being programmers, a step toward that abundance in which building stops being a technical privilege.