On stage at TechCrunch Disrupt on October 16, Replit CEO Amjad Masad said his company no longer works with Anthropic. Replit is one of the best-known platforms for building apps straight from the browser, and a good share of its recent growth was built on Claude models.
Cost, partnership, performance
Masad gives three reasons in one sentence: “It was a decision based on both cost, partnership, and performance.” Just before it comes the line that travelled: “I think ChatGPT is better now.”
He points to measurements. Replit runs an internal benchmark and publishes a public one too. On both, Masad says, Sol and Astra do better at building vibe-coded applications from scratch. Replit had already announced an OpenAI partnership back in August, though nothing then suggested a switch.
He also gave numbers on the business: enterprise was 2% of revenue last year and is 15% this year, and more than half of revenue now comes from B2B.
The pushback came the same day
TechCrunch quotes a person familiar with the partnership, speaking anonymously, who says it is still in place. There are traces supporting that. A Replit blog post from late September names Claude Fable 5 and Claude Fable 5.1 as models in the Replit Agent production environment. Around the same time a Replit product employee posted on X about Opus 5.5 and praised what it could do. TechCrunch has asked Replit for clarification and is waiting.
So who is right? Both can be true at once. A contract can run on while the models on the default path have already changed. And a CEO on a conference stage is talking about direction, not about what legal has filed.
Models have become interchangeable
The interesting part is not who leads today. It is how casually the swap happens. Two years ago, the model behind a coding platform was a strategic commitment. Now it is a procurement line settled by a benchmark and a price list.
For Anthropic that stings twice. Platforms are visible, and what they run shapes the impression of developers who never compare for themselves. And Masad names cost first. That is where the fight actually is, not in the benchmarks.
What I miss are numbers. “Better on our benchmark” says little until someone states what was measured and at which settings. The public benchmark exists, and I will look at it before I believe the sentence.