Replit Quietly Swapped Claude for ChatGPT — and the Reasons Are Blunt

If you build apps with AI inside Replit, the model doing the heavy lifting may have quietly changed under your feet. At TechCrunch Disrupt this week, Replit CEO Amjad Masad said the company is no longer working with Anthropic as its primary model partner — and his explanation came down to three very practical words: cost, performance, and partnership.

The short version is that OpenAI's ChatGPT is now the default engine behind Replit Agent. Claude is still on the menu, but the default matters more than most people realize, because most users never touch the model picker.

What Masad actually said on stage

According to TechCrunch, Masad told the Disrupt audience that Replit's decision wasn't emotional or political. He pointed to an internal benchmark the company uses to measure how well models build "vibe-coded" applications from scratch, plus a public benchmark they publish. On both, he said OpenAI's Sol and Astra models came out ahead. His blunt summary: "I think ChatGPT is better now."

He framed the switch as a business call, not a breakup. Replit announced its OpenAI partnership back in August, and Masad described OpenAI as the preferred partner for powering Replit Agent on performance, economics, and the strength of the relationship.

The nuance: Claude didn't disappear

This is where the story gets more interesting than a simple migration headline. Masad clarified after his talk that customers can still select Anthropic's Claude and other supported models inside Replit. The company describes itself as model agnostic, and that's probably the right posture for a platform whose users have strong opinions about which model writes better code.

Still, defaults drive behavior. If ChatGPT is the out-of-the-box choice, Anthropic's share of Replit's inference bill will shrink even without an official termination. Until recently, Masad said Replit had been one of Anthropic's top-five customers — a detail that shows just how much revenue is at stake when a single platform flips its default.

Not everyone agrees with the framing

TechCrunch reported that people familiar with the relationship pushed back on Masad's characterization, with one source saying the partnership was still in place. There's public evidence backing that skepticism: a late-September Replit blog post referenced Anthropic's Claude Fable 5 and Fable 5.1 running in Replit Agent's production environment, and a Replit product employee was openly praising Opus 5.5 on X around the same time.

So the honest reading is messier than either side's talking points. A default can change while a commercial relationship technically survives. Both things can be true, and for developers the practical outcome is the same: the default model in your workspace is now ChatGPT.

The growth story behind the switch

Model choice aside, Masad shared numbers that explain why a company like Replit can afford to be picky. Enterprise was roughly 2% of revenue last year. This year he puts it near 15%, with more than half of all revenue coming from B2B contracts. He told TechCrunch he expects enterprise and government customers to account for the majority of revenue in 2027.

He also made a point about AI safety that stuck with me. Agents, he argued, need a habitat: orchestration, error recovery, sandboxing, networking, identity, and authorization. Get that infrastructure right, he said, and you can deploy AI safely. That's a far more useful framing than treating safety as a filter bolted on at the end.

Why this matters for builders

If you're shipping on a managed AI coding platform, you don't fully control your stack. The model underneath you can shift because of a pricing negotiation or a benchmark result, and your prompts, costs, and output quality can move with it. The lesson isn't to avoid these platforms — it's to keep your own evaluation harness handy so you can tell when a silent default change starts producing worse code.

Model loyalty is a marketing story. Benchmarks, invoices, and shipped features are the real scoreboard.