Replit says it has ended its Anthropic partnership despite pushback from insiders

Replit says it has ended its Anthropic relationship

Replit is no longer working with Anthropic, according to the AI coding platform's CEO and co-founder Amjad Masad, who made the announcement while speaking at TechCrunch Disrupt on Thursday.

According to TechCrunch, Masad said the decision came down to "cost, partnership, and performance," and he was blunt about how the two labs' models compare: "I think ChatGPT is better now."

Replit had already announced a partnership with OpenAI in August, but as TechCrunch notes, nothing at that point suggested the Anthropic relationship was over. Masad said Replit assesses models with an internal benchmark as well as a public one, and claimed that two models he named, Sol and Astra, perform better at building "vibe-coded" applications from scratch. For those reasons, he said, the company felt OpenAI was "a better partner for now."

A split that not everyone agrees happened

The announcement was quickly contradicted. TechCrunch reports that one person familiar with the Anthropic–Replit arrangement, who asked not to be named, said the partnership is in fact still in place.

There is also evidence that Anthropic's models remained in production very recently: a Replit blog post published in September cited Anthropic's Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent's production environment. TechCrunch says it went back to Replit for clarification and will update its story if it hears back.

The stakes of the disagreement are real. Replit's platform — a browser-based environment where users build and deploy applications — ran on Anthropic's Claude models under the partnership, and TechCrunch reports that arrangement was credited with helping the company significantly grow its revenue.

Business customers now drive most of Replit's revenue

On the same stage, Masad described how Replit's revenue mix has shifted. While declining to share ARR figures, he said enterprise contracts went from roughly 2% of revenue last year to around 15% this year, and that more than half of the company's revenue now comes from B2B business.

He also argued that safe AI deployment depends on building the right environment for agents. Just as humans need an office, equipment and comfortable conditions to do good work, Masad said, agents need a "habitat" with good orchestration, error recovery, security and sandboxing, networking, and identity and authorization controls — the pieces he says must be figured out before AI can be deployed safely.

Why it matters

AI coding platforms are among the largest consumers of frontier models, so when one switches suppliers, significant inference volume and revenue move with it. For Anthropic, whose Claude models helped power Replit's growth, losing a flagship coding customer to OpenAI would be a notable blow. For OpenAI, it would be further evidence of momentum among developers.

The episode also shows how fluid these arrangements have become. Masad's phrasing — a better partner "for now" — suggests that platform companies increasingly treat model providers as interchangeable, swapping them as benchmarks and unit economics shift.

Finally, the contradictory accounts underline how opaque these partnerships are. Developers learned which models were running their agents from a CEO's on-stage comments and a months-old blog post rather than a formal announcement, and even now it is unclear whether the switch is complete. Until Replit clarifies, both the timeline and the current state of the Anthropic relationship remain in question.