A Single "We No Longer Work With Anthropic" Walked Back Over the Weekend — Replit's CEO on Switching to OpenAI Onstage, and the Counterclaims From People Close to the Deal

At TechCrunch Disrupt in mid-October, Amjad Masad, CEO of the AI coding platform Replit, said onstage that "Replit is no longer working with Anthropic," stating that the company had switched the default model provider behind Replit Agent to OpenAI. A few days later, however, Masad walked the remark back on X, saying "Replit remains model-agnostic." On top of that, a person said to be familiar with the partnership told TechCrunch that "the partnership is still ongoing." Here we separate out what was said, how, and what still cannot be confirmed.

"No Longer Working Together" — The Reasons for the Switch, Laid Out Onstage

Masad gave roughly three reasons for the switch. Rather than letting them blur into prose, here they are, organized:

• Cost: OpenAI's per-token pricing is lower.
• Partnership terms: The commercial terms as a partner were more favorable.
• Performance: In Replit's internal benchmarks, OpenAI's models (Sol and Astra) scored high for "building an app from scratch." He was reported to have added that "ChatGPT is better right now."

Masad also said that Replit had once been one of Anthropic's "top five customers." If true, the switch is less a single user's preference than a matter of sizable orders on the move. That said, this is his own statement, not a figure disclosed by either company — a point worth keeping in mind.

The Assertion That Softened Over the Weekend

As the remark drew attention, Masad softened his wording on X over the weekend. In his words: "Replit remains model-agnostic, and users can still choose Anthropic's Claude or other supported models. OpenAI has simply become the 'preferred partner.'" In other words, the flat assertion that "we no longer work together" was replaced by a far cooler story: "we've leaned Replit Agent's default toward OpenAI / Claude remains available as an option."

Traces of Claude Still in Production

Counterevidence has also emerged around the actual state of the partnership. One person told TechCrunch that "the partnership is still ongoing," and Replit's own recent messaging was reported to appear to back this up. Specifically, in a Replit engineering blog post dated September 29 (explaining model routing), OpenAI's GPT-6 Astra was front and center, yet Claude Fable 5 and Fable 5.1 were listed as models running in production Replit Agent. Neither company has issued an official statement. How far "no longer working together" extends in practice, whether Claude has actually been removed, and if so when, cannot be confirmed at this point. This article will not assert otherwise.

The "Side That Picks the Model" Is Starting to Hold Bargaining Power

Setting aside whether the remarks are true, the episode clearly brought one thing into focus: coding platforms have begun to wield bargaining power as the "side that chooses" the provider of the models underneath them. At the end of August, OpenAI cut off GPT access for Cursor, showing that the supply tap can become a competitive weapon. This time is the flip side: the platform publicly declares "which model to set as the default," weighing cost against performance and dangling the threat of a switch.

Behind this is a shift in Replit's own business. The company's revenue from enterprise reportedly grew from about 2% a year earlier to roughly 15%, with B2B coming to account for the majority. The more its customers skew toward businesses, the more directly cost (i.e., model token fees) and output quality feed into the P&L. It is natural to see the choice of provider gaining weight — moving from "preference" to "procurement decision."

Default Models Can Be Swapped Without Your Consent

A note on what this implies for the user side as well: a platform's "default model" can be swapped out behind the scenes, without the individual user's consent. If the model your app rests on changes, so do its output quirks, its token pricing, and sometimes the quality of the results. Once you build on top of a platform, this is a risk you should, to some degree, factor in from the start.

One more point worth watching: the performance comparison cited as grounds for the switch came from the vendor's own internal benchmarks, not a third party. A claim of "superior by our own measurement" — whoever makes it — is something the audience should not take at face value. As in this case, there are examples where the principal's own statement softens within days and people close to the matter push back. Don't rush to assert; back things up with primary sources and real-world behavior. The same caution, it seems, is required of those choosing a coding platform.

References: TechCrunch / Unite.AI / explainX / BigGo Finance