Replit said it was leaving Anthropic and walked it back within 24 hours: a change of default model, not a divorce

By Tech Insider · October 19, 2026.

On October 16, at TechCrunch Disrupt, Replit CEO Amjad Masad said his company was no longer working with Anthropic. According to TechCrunch, he cited "cost, partnership and performance." He added that "ChatGPT is better now" and that Replit's internal benchmarks favored OpenAI for building apps from scratch.

That account didn't last. According to the same outlet, a person familiar with the relationship maintained that the partnership was still active. Masad clarified that Replit "remains model-agnostic": users can choose Claude or other models, and OpenAI becomes the "preferred partner" for Replit Agent, the platform's code-generating agent.

There is also a documentary record. On September 29, Replit published "Free the models: Harness design at the frontier" on its engineering blog. The post reportedly compared production traffic from Claude Fable 5 (August), Claude Fable 5.1 (September) and GPT-6 Astra (September). In other words, Claude was still in production weeks before the announcement. That doesn't prove the contract was still in force in October, but it does clash with the original statement.

It's worth noting what isn't known. There are no figures on usage share, token volume (the units of text a model bills for), per-model cost or revenue impact. As of October 19, neither Anthropic nor OpenAI had commented. Replit's internal benchmark results are also the company's own claims, not verified by third parties.

**Our take.** This is a change of default provider framed as a breakup. And that says more about the industry than about Replit. In general, AI coding platforms such as Cursor, Windsurf and GitHub Copilot already route requests across multiple models, because no lab leads on cost, latency and quality all at once. The model is becoming a commodity input; value is shifting to the orchestrator, the agent and integration into the development environment.

That's why the headline "Replit drops Anthropic" was appealing and misleading: it confused a routing decision with a strategic one. OpenAI wins free publicity; Anthropic takes a small reputational hit, not a financial one, as far as we know; and Replit pays the biggest price, because its CEO was contradicted by his own engineering blog. For an enterprise customer, the consistency of a vendor's messaging is also part of the product.

There's a practical lesson for anyone building on these platforms: verify which model is actually running in your workspace and log it. A change in default model alters cost, behavior and compliance requirements without any major notice.

And a deeper one, in line with our thesis: that the model frontier is so contested and so interchangeable is good news in the long run. Competition drives prices down and spreads power around, and that brings abundance closer. In the short term, however, it forces us to read the marketing with more skepticism: between the stage and the technical blog, the blog usually wins.