Replit CEO Says the Company Has Stopped Working With Anthropic

Table of Contents

• You might want to know
• Replit’s Relationship With Anthropic
• Key Insights Table
• Afterwards...

You might want to know

• Why did Replit’s CEO say the company had stopped working with Anthropic?
• Why did people familiar with the partnership question whether the relationship had actually ended?

Replit’s Relationship With Anthropic

Replit, an AI-assisted coding platform used by startup founders and developers, is no longer working with Anthropic, according to Replit CEO and co-founder Amjad Masad. He made the statement while speaking onstage at TechCrunch Disrupt on Thursday. Masad said the decision reflected several considerations, including cost, performance, and the companies’ partnership arrangements. His comments indicated that Replit had selected another provider as a better fit for its needs at this time.

The announcement attracted attention in part because Replit had previously disclosed a partnership with OpenAI in August, but that news had not been accompanied by any indication that its relationship with Anthropic had ended. Replit had used models from Anthropic’s AI lab, including Claude, as part of the technology behind its platform. The difference between the earlier public announcements and Masad’s new remarks left questions about when any change took place and how broadly it applied.

Asked about the decision, Masad said he believed ChatGPT had become better. He explained that Replit uses an internal benchmark to assess measures such as performance and also publishes a public benchmark. According to Masad, the results consistently showed that Sol and Astra performed better at building “vibe-coded” applications from scratch. He presented the choice as one based on more than a single benchmark result: cost, the partnership itself, and performance all contributed to Replit’s assessment.

Masad described the move as a decision made for the present, saying Replit felt the alternative provider was a better partner “for now.” That qualification is notable. It frames the change as a current business and product decision rather than a definitive statement about what Replit might choose in the future. It also reflects how AI companies may reassess model providers as products, pricing, and technical capabilities change.

Replit offers a cloud-based software development platform that allows users to build and deploy applications directly in a web browser. Rather than requiring users to manage a local development environment, the service brings development tools into an online workspace. AI models can support this experience by helping users generate, edit, and troubleshoot code. The choice of models and the supporting infrastructure can therefore influence both the platform’s functionality and its operating costs.

Under the earlier arrangement, Replit used Anthropic’s Claude models to power elements of its platform. The partnership was reported to have helped Replit grow its revenue significantly. That background makes the CEO’s remarks consequential: the relationship was not simply a matter of using an interchangeable tool, but one connected to Replit’s product offering and commercial development. Still, the available information does not explain the precise timing of any transition, whether every Claude-powered feature was removed, or what technical changes were required.

Some people familiar with the relationship pushed back on the suggestion that the partnership had ended. One person who knew about the arrangement, and who requested anonymity, said it was still in place. That account conflicts with Masad’s description and leaves the status of the relationship unclear based on the statements available. The disagreement may reflect different understandings of what it means for a partnership to be “in place,” or it may indicate that the companies’ arrangements were changing while some uses continued.

There is also evidence that Replit was using Anthropic products as recently as September. A blog post published that month identified Anthropic’s Claude Fable 5 and Claude Fable 5.1 as models running in the production environment for Replit Agent. This reference does not, by itself, establish whether the partnership remained active at the time of Masad’s comments. It does, however, provide a recent public example of Anthropic models being associated with a live Replit product, making the timeline an important part of the story.

TechCrunch said it had contacted Replit again to seek clarification and would update its report if it received a response. Until the company provides further detail, several aspects remain unsettled: the exact scope of the stated separation, whether any contractual or technical relationship continued, and how the use of models in Replit Agent relates to the CEO’s comments. The conflicting accounts make it useful to distinguish between what Masad said publicly, what an unnamed source said about the partnership, and what the September blog post documented.

At the same event, Masad also discussed Replit’s changing revenue mix. He said enterprise customers were becoming a larger part of the business, while declining to provide annual recurring revenue figures. He described enterprise as accounting for 2% of revenue last year and said it represented 15% of revenue this year. He also stated that more than 50% of Replit’s revenue was B2B. These figures suggest that Replit is seeking growth beyond individual developers and small teams, although Masad did not disclose an ARR total.

Enterprise growth can raise expectations for reliability, security, and administrative controls. Companies adopting AI-assisted coding tools may need confidence not only in code-generation quality, but also in how the system handles access, data, networking, and failures. Masad’s remarks about enterprise revenue and AI safety therefore connect two parts of Replit’s strategy: expanding its customer base and building an environment in which AI agents can operate more safely and reliably.

Masad described AI agents as needing a suitable “habitat” to perform useful work. He compared the requirements of agents with the conditions people need in a workplace, such as an office, food, air-conditioning, and a computer. For AI agents, he said, that environment includes orchestration, error recovery, safety and security, sandboxing, networking, identity, and authorization. His point was that deploying capable models is not enough on its own: the surrounding systems determine whether agents can work effectively and within appropriate safeguards.

That argument places infrastructure at the center of AI safety. Orchestration can help coordinate actions, while error recovery can prevent a single failure from derailing a task. Sandboxing and security controls can limit what an agent is able to access, and identity and authorization systems can establish which actions are permitted. Networking also matters when an agent must interact with tools or external services. In Masad’s view, these elements together create the conditions for safer deployment.

The central unresolved point is whether Replit’s collaboration with Anthropic had fully ended when Masad spoke: his statement said it had, while an unnamed person familiar with the arrangement said it remained in place. The September references to Claude models add context, but do not independently settle the status. More information from Replit would be needed to establish the timing and scope of any change.

Masad identified cost, partnership considerations, and performance as the reasons behind Replit’s decision, and said the company’s benchmarks favored Sol and Astra for building vibe-coded applications from scratch. His comments provide a view into how Replit evaluates model providers, while the contradictory accounts underscore how partnerships can be difficult to characterize from outside the companies involved.

Key Insights Table

Afterwards...

Replit’s next steps will be shaped by how clearly it explains the status of its relationship with Anthropic and how it describes the models powering its products. The company’s public comments, the unnamed source’s account, and the September product reference point to different parts of a timeline that has not yet been fully clarified. Any follow-up from Replit could help establish whether the change was complete, gradual, or limited to particular products or uses.

The broader business story will also depend on Replit’s ability to serve enterprise customers while maintaining a secure and dependable platform. Masad’s revenue figures indicate that business customers have become a more significant part of the company’s mix, while his comments about agent infrastructure highlight the operational requirements that come with wider AI deployment. For users and customers, the most important developments will be the quality of the tools, the reliability of the underlying systems, and the safeguards that govern what AI agents can do.

For now, the account offers a snapshot of a fast-moving market in which model performance, cost, partnerships, and safety infrastructure all influence product decisions. Replit’s choice may evolve as these factors change, but the company’s emphasis on evaluation and agent environments suggests that choosing a model provider is only one part of building an effective AI development platform.