Replit's CEO Says the Company Has Dropped Anthropic for OpenAI — a Claim Insiders Dispute

Replit has stopped working with Anthropic, according to its own chief executive — but the split itself is already contested. Amjad Masad, the coding platform's co-founder and CEO, said onstage at TechCrunch Disrupt on Thursday that the company that once ran its agent on Claude models is now leaning on OpenAI instead.

"I think ChatGPT is better now," Masad said, framing the decision as a pragmatic one based on cost, the quality of the partnership, and model performance. He was explicit that the choice is provisional: OpenAI is the better partner "for now." Replit first announced its OpenAI partnership in August, but at the time there was no indication that the Anthropic relationship had ended.

The performance argument rests on Replit's benchmarks. Masad said the company uses an internal benchmark to evaluate models and also publishes a public one, and that OpenAI's Sol and Astra consistently do better at building so-called vibe-coded applications from scratch — the core flow Replit Agent sells. He added that Replit's strategy is to assign different models to different jobs, chasing the highest quality at the lowest cost rather than standardizing on one supplier.

There is a wrinkle, though. A person familiar with the relationship told TechCrunch the Anthropic partnership remains in place, pushing back on Masad's characterization. And Replit's own records complicate the story: a September 29 engineering post titled "Free the models: Harness design at the frontier" compared a week of production traffic on Claude Fable 5 in August with Fable 5.1 and OpenAI's GPT-6 Astra in September — confirming Claude was still in production just weeks before the Disrupt appearance.

The commercial backdrop explains the leverage shift. Masad said enterprise contracts grew from about 2 percent of revenue last year to roughly 15 percent this year, with more than half of revenue now business-to-business. A customer of that profile shopping between labs on price per token and narrow-task benchmarks is exactly the kind of buyer that keeps pricing pressure on both OpenAI and Anthropic.

Masad also used the stage to lay out what he thinks safe agent deployment requires: a "habitat" with orchestration, error recovery, security, sandboxing, networking, and identity and authorization. Once that scaffolding exists, he argued, AI agents can be deployed with confidence — a notable position from the CEO of a platform whose own agents have had public incidents in the past.

For Anthropic, losing a visible developer-tools customer to OpenAI — even a disputed one — is the kind of signal that tends to surface quickly in pricing changes or new benchmark claims. For every startup wiring its product tightly to one model provider, the episode is a reminder that model partnerships are infrastructure, not marriage: cheaper to swap than the roadmap assumes.