Replit drops Anthropic as default for ChatGPT, CEO says
Replit's CEO used the TechCrunch Disrupt stage to say what the company had never said out loud before: the Anthropic partnership is over. The clarification that followed matters as much as the statement — and the two don't entirely agree.
Replit CEO Amjad Masad said the AI coding platform is no longer working with Anthropic, citing cost and telling the Disrupt audience plainly: "I think ChatGPT is better now." Speaking onstage Thursday, Masad said the decision rested on three legs — cost, partnership, and performance — and pointed to benchmark results rather than sentiment. "We publish a public benchmark as well, and consistently, Sol and Astra perform better on building vibe-coded applications from scratch," he said, as TechCrunch reported from the stage. "It was a decision based on both cost, partnership, and performance. And for all these reasons, we just felt like they were a better partner for now." The framing matters: Replit struck a deal with OpenAI in August, and until Thursday there was no public sign that the Anthropic relationship had cooled. Masad noted that Replit had been, until recently, a top-five customer of Anthropic's — which is exactly why a switch moves money. A top-five customer that quietly reroutes most of its default traffic to a rival is not a rounding error for either lab.
The follow-up clarification is where the story gets interesting. After his onstage remarks, Masad told TechCrunch that customers can still access Anthropic's Claude models on the platform, but OpenAI's ChatGPT is now the default for Replit Agent. In a written statement he put it more carefully: "Replit remains model agnostic. Our users can choose Anthropic's Claude or other supported models… OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership." Because most users never change the default, that shift alone could cut Replit's spend with Anthropic substantially — the default is the product for the long tail of developers who never open a model picker.
Not everyone connected to the partnership agrees the partnership ended. People familiar with the arrangement told TechCrunch that it is still in place, and there is documentary evidence on Replit's own properties to back them up: a late-September blog post cited Anthropic's Claude Fable 5 and Fable 5.1 as models running in Replit Agent's production environment, and around the same time one of Replit's product employees was on X promoting Opus 5.5's capabilities. RuntimeWire, independently reporting the same remarks, flagged that Replit's September production notes still covered Claude models in the Agent environment, and that Replit launched an integration inside Claude in June — leaving the scope of the break unclear. Both readings can be true at once: "no longer working with Anthropic" may describe the commercial default and roadmap direction, while the technical integration and existing contract run on for now. The honest summary is that Replit has demoted Anthropic from preferred partner to optional model, and is letting a contradiction between its CEO's stage remarks and its insiders' account stand in public.
Why it matters beyond one platform: Replit is a bellwether for how coding agents get bought. It is not choosing a model the way a researcher chooses a model — it is choosing which lab gets the default slot in a product where most users accept whatever ships. The August OpenAI deal was the warning shot; Thursday was the confirmation that default placement, not raw capability alone, is the real contest between labs in the coding-tools market.
That contest has a clear economic logic. Model APIs are priced per token, agent sessions run long, and coding agents burn tokens faster than chat use does — a platform that routes millions of agent sessions through one provider is making a procurement decision at scale, not an editorial one. Masad's mention of cost alongside performance is the tell: even if Claude and GPT-class models were tied on capability, the default slot is where a company feels pricing pressure first. For Anthropic, losing Replit's default is not only lost inference revenue; it is losing the surface where developers form habits. Developers who start every project inside Replit Agent learn one model's quirks and failure modes first, and those habits carry into the tools they choose next. OpenAI, meanwhile, gets what its August partnership promised — a generation of builders whose first agent experience runs on its models by default.
The financial backdrop raises the stakes: Masad said enterprise went from 2% of revenue last year to roughly 15% this year, more than half of revenue is now B2B, and he expects enterprise and government customers to make up more than half of Replit's revenue in 2027. Enterprise buyers ask which model powers the agent by default; Replit's answer just changed. Masad also argued onstage that agents need a "habitat" — orchestration, error recovery, sandboxing, identity and authorization — to work safely, which is Replit's pitch for why the platform, not the model, is the layer worth paying for. It is a coherent story: if the platform is where the safety work lives, the platform can afford to be model-agnostic and let benchmarks and invoices decide. We have seen the other side of this customer-flux story before — Pentagon says it has stopped using all Anthropic products.
What to watch: whether Anthropic responds publicly, whether Replit's public benchmark methodology gets scrutinized now that it is doing load-bearing work in a vendor decision, and whether Claude quietly disappears from Replit's production notes this quarter — the documents will settle what the stage remarks and the insider account cannot.
If you build on Replit, does the default model decide what you ship — or do you pick your model every time? Tell us in the comments.
