Replit's CEO Said He Dropped Anthropic. His Own Company Wasn't So Sure.
On stage at TechCrunch Disrupt on October 15, Replit CEO Amjad Masad told the audience his company was "no longer working with Anthropic," citing cost and performance, and said flatly that "ChatGPT is better now" for the vibe-coded apps Replit's agent builds. It was a clean, quotable line about an AI-native startup walking away from one of the two labs its product was built on.
It didn't hold up. Within days, TechCrunch reported that a person familiar with the Anthropic partnership said it was still in place, and that Claude models — Fable 5.1, Opus 5.5 — had been visibly running in Replit's production environment as recently as late September, weeks before Masad's remarks. Replit's own blog and social accounts had been promoting Claude-powered features in that same window. By the following Monday, the company's position had softened considerably: OpenAI is now the "preferred partner" and default model for Replit Agent, not Anthropic's replacement. Users can still choose Claude. Masad said Replit "remains model agnostic."
A breakup announced before it happened
That gap between the stage and the server logs matters more than which model Replit prefers this month. Masad made the cost argument at a moment when Replit has a strong incentive to look disciplined about AI spending to a very different audience than it had a year ago: he told the Disrupt crowd that enterprise and government revenue is on track to grow from roughly 2% of the business last year to about 15% this year, with more than half expected within two years. A public story about walking away from an expensive vendor over cost and performance plays well to that buyer. Whether the underlying contract, discount terms, or actual production traffic moved at all is a separate question Replit has not answered with numbers.
The buy side is doing the same thing, more quietly
Replit's non-breakup landed in the same two weeks that Accenture told employees it would centrally fund only $29 a month of their Claude usage, after some had been running bills of $500 to $1,000. The internal memo, first reported by 404 Media, said "providing a WBS and obtaining a license does not guarantee indefinite or ongoing access to Claude" — blunt language for a firm with roughly 799,000 employees. Accenture's own research has it telling clients that 82% of executives are increasing AI investment, which made the internal memo read, to several outlets, as a company constraining its own staff's AI use while selling the opposite story to everyone else.
Neither event changes much about what AI actually costs to run. What they share is a pattern: both Replit and Accenture reached for a dramatic public statement about an AI vendor relationship — a severed partnership, a hard dollar ceiling — that turned out, on inspection, to be softer than advertised. Replit still serves Claude on request. Accenture's cap is a budget line employees can route around by charging a project code, not a block. The theater is doing real work. As model providers keep cutting list prices while usage keeps outrunning those cuts, enterprise buyers are learning that a loud complaint about a vendor is cheaper leverage than a renegotiated contract, and a lot easier to quietly walk back later.