Replit CEO says it stopped working with Anthropic, but an insider disputes it

What Masad said on stage

Replit's chief executive Amjad Masad said onstage at TechCrunch Disrupt that the AI coding platform is no longer working with Anthropic, according to TechCrunch, which reported the remarks on 16 October 2026. Replit, a browser-based platform where people build and deploy apps by describing them to an AI agent, has relied on Anthropic's Claude models for much of its growth.

Masad gave three reasons: cost, partnership and performance. "I think ChatGPT is better now," he said, according to TechCrunch. He said Replit uses an internal benchmark and also publishes a public one, and that "consistently, Sol and Astra perform better on building vibe-coded applications from scratch." Sol and Astra are OpenAI models. "It was a decision based on both cost, partnership, and performance," he said, adding that Replit felt OpenAI was "a better partner for now."

Masad also shared rare numbers about the business. He said Replit is not yet sharing annual recurring revenue, but that enterprise contracts went from about 2% of revenue last year to about 15% this year, and that more than half of revenue now comes from business customers.

Why the claim was disputed within hours

The statement did not go unchallenged. TechCrunch reported that people familiar with the relationship rejected Masad's account, and that one person, who asked not to be named, said the partnership between Replit and Anthropic was still in place. TechCrunch said it had asked Replit for clarification. AI Lookout found no public statement from Anthropic on the remarks as of 17 October.

That leaves two readings. Masad may mean that Replit has stopped using Claude as its default model, or ended some commercial arrangement, while still calling Anthropic's API for some tasks. Or Replit may be winding Claude out completely. RuntimeWire summed up the gap: Replit has not said what "stopped working" covers.

What Replit's own records show

Replit's own publications suggest Claude was in production very recently. A Replit engineering post titled Free the models, published on 29 September 2026, describes how its agent's core loop decides when to hand work to smaller subagents. It compares three models running in Replit Agent production: Anthropic's Claude Fable 5 in August 2026, and Claude Fable 5.1 and OpenAI's GPT-6 Astra in September, one week of traffic per model.

The post is complimentary about both companies. It says the code quality of Astra and Fable 5.1 let Replit use its code-review subagent less with no drop in its evaluation scores, and that it had not seen that level of engineering quality from any model before. It also records how differently the models behave. At medium reasoning effort, Fable 5.1 dispatched a subagent on 21% of turns and handed work to a general worker on 2.3%, while Astra did so on 36% and 20%. Astra returned to an already briefed subagent on 42% of dispatches, against 29% for Fable 5.1.

The headline benchmark results in the same post use Astra as the core model in Replit's highest-quality Max mode. On DeepSWE v1.1 Replit reports 72% at $2.11 per task, and on Terminal-Bench 4.0 49% at $2.53 per task, both Replit's own runs. Those numbers fit Masad's argument that OpenAI models now lead for Replit, but they do not show that Claude was removed. TechCrunch also noted a Replit product employee posting about using Claude Opus 5.5 around the same time.

How Replit got closer to OpenAI

The tilt toward OpenAI has been public for two months. On 19 August, OpenAI published a customer story saying GPT-5.6 Luna powers Replit's new Free Mode, which lets people plan and explore ideas with the agent without using paid credits. Harder tasks can be routed to GPT-5.6 Sol. Masad is quoted thanking OpenAI for "the price cuts that you made recently," which he said made it possible to offer Free Mode to millions of users.

TechCrunch called that August deal a partnership but noted there had been no sign Replit had left Anthropic. Taken together with the September post, the picture is a company moving more of its traffic to OpenAI over the summer while still testing and running Claude, and a CEO now describing that shift in much stronger terms.

Why model choice is now a business decision

For coding platforms, picking a model is no longer just a technical choice. Inference is one of their biggest costs, and a free tier only works if tokens are cheap. Masad's first reason was cost, and the OpenAI story ties Free Mode directly to OpenAI's price cuts.

The model makers are also competitors. Anthropic sells Claude Code and has been expanding it for teams, including general availability of Claude Code Review for Team and Enterprise plans and a report on how its own engineers doubled merged pull requests with AI review. OpenAI has Codex. Platforms like Replit sit on top of those same labs while selling to some of the same users, which makes multi-model flexibility a survival strategy. Crypto Briefing reported that Masad stressed at Disrupt that Replit does not want to depend on any single model provider.

Masad also talked about safety, saying agents need a "habitat" with good orchestration, error recovery, sandboxing, networking, identity and authorization before they can be deployed safely. That is a pitch for Replit's own infrastructure: the platform, not the model, is what Replit wants customers to buy.

What it means for Replit users

Replit has not announced any change to the models users can pick, and there is no public notice that Claude options are going away. Users who care which model writes their code should check the model settings in their projects and watch Replit's changelog and blog for a formal statement.

For businesses that signed enterprise contracts, the practical questions are data handling and continuity. If a contract names specific model providers, a supplier change could matter for compliance reviews, so it is worth asking Replit directly which providers process their code today.

What is still unclear

The scope of the split is the main open question. Replit has not published a statement, and Anthropic has not answered Masad publicly. We do not know whether Replit has ended a commercial agreement, stopped defaulting to Claude, or removed Claude entirely, or when any change took effect.

We also have not seen the public benchmark Masad mentioned in a form that compares the latest Claude and OpenAI models on building apps from scratch, so his performance claim rests on Replit's own testing. We will update this article if Replit or Anthropic clarify the relationship.