Replit says it ended work with Anthropic; insiders dispute the split

On stage at TechCrunch Disrupt on Thursday, Replit CEO and co-founder Amjad Masad said the company is no longer working with Anthropic. Replit is the cloud-based development platform where users build and deploy apps from inside a browser, and under its Anthropic partnership it ran its platform on the lab's Claude models. A partnership that, according to TechCrunch, helped the company grow its revenue significantly.

The reasons he gave: cost, partnership, performance

Masad offered three reasons, in that order of emphasis in his own words. On performance, he said: "I think ChatGPT is better now." Replit runs an internal benchmark to assess metrics like performance and publishes a public one as well. In his account, OpenAI's Sol and Astra "consistently perform better on building vibe-coded applications from scratch." The split, he said, was "a decision based on both cost, partnership, and performance," and for those reasons the OpenAI side was the better partner "for now."

The "for now" matters. Replit announced its partnership with OpenAI back in August, and there was no public sign before Thursday that the Anthropic relationship had ended. If anything, the public record pointed the other way.

The part that is disputed

Masad's comments were rejected by people familiar with the partnership. One insider, speaking on condition of anonymity, told TechCrunch the partnership was still in place.

There is also a paper trail that complicates the clean break. TechCrunch points to Replit's own website, blog posts and social media showing Anthropic products in use as recently as a few weeks before the stage appearance. A late September blog post names Claude Fable 5 and Claude Fable 5.1 as models running in Replit Agent's production environment. Around the same time, a Replit product employee posted on X promoting the capabilities of Anthropic's flagship model Opus 5.5.

TechCrunch reached back out to Replit for clarification and updated the story with that context. As of publication, the company had not confirmed the end of the partnership.

What he did confirm: the enterprise shift

Away from the model question, Masad shared the clearest numbers the company has put out on its revenue mix. Enterprise was roughly 2% of Replit's revenue last year. This year, enterprise contracts account for around 15%, and more than 50% of total revenue is B2B. The company is not sharing ARR figures yet.

He also made a point about agent infrastructure that is worth keeping on file. In his framing, agents need a "habitat" to do good work: good orchestration, error recovery, safety and security, sandboxing, networking, and identity and authorization. "Once you figure all of that out, then you can deploy AI safely," he said.

What this is worth, as a data point

One CEO's stage announcement against one insider's denial does not settle who is right, and the article itself is still pending Replit's response. What the episode does establish is more useful: a production application platform of Replit's size can route its core agent product between frontier labs on the basis of cost and benchmark performance, and it will say so publicly mid-contract.

If you build on a single lab's API, that is a concrete reminder that model routing is a decision your platform vendor makes for you, that those decisions now turn on cost and evals rather than loyalty, and that "the better partner for now" is the honest name of every model contract.