Replit CEO and co-founder Amjad Masad says the AI coding platform is no longer working with Anthropic and has chosen a different direction based on cost, partnership and performance. Yet people familiar with the relationship disputed his account, while recent Replit material still identified Anthropic models in production, according to TechCrunch. The supplier decision is therefore significant, but its operational scope remains unresolved.
For SME teams using AI to build internal commercial applications, the distinction matters. A provider announcement does not reveal which model handles a particular request, whether an existing application has moved, or how the change affects the finished workflow.
Cost, partnership and app performance drove the decision
Masad said Replit considered several factors, including cost. He also said ChatGPT was now better and described OpenAI as the better partner “for now.” Replit had announced an OpenAI partnership in August without disclosing that its work with Anthropic had ended, TechCrunch reported.
Performance was presented as another part of the mechanism. Masad said Replit uses an internal benchmark to assess measures such as performance and also publishes a public benchmark. He said Sol and Astra consistently performed better when building vibe-coded applications from scratch, although the report does not provide the benchmark methodology, individual results or a direct breakdown of each model’s contribution.
This is not simply a comparison between two chat interfaces. Replit provides a browser-based environment in which users can develop and deploy applications. The underlying model operates inside a broader product that also shapes prompts, tool use, code execution and deployment. A decision attributed to model performance therefore reflects the application-building system, not necessarily a stand-alone model tested in isolation.
Conceptual illustration of the factors Replit says informed its model-provider decision.
An upstream switch is not a customer migration
Replit previously used Anthropic’s Claude models to run its platform, according to TechCrunch’s account of the relationship. The report does not specify which Replit functions have now changed provider, whether the change covers every user or how existing projects are handled.
That missing detail defines the practical question for a commercial operations team. The relevant unit of evaluation is the application workflow delivered through Replit: what the agent builds, how reliably it edits an existing project, whether the result deploys correctly and how much intervention the task requires. The identity of the upstream model is one input into that experience rather than the entire product.
Consider a hypothetical team building an internal quote-request portal. A successful initial interface would not establish that the agent can preserve existing business rules, repair later changes or maintain the application after deployment. Those stages require their own tests even when a platform reports stronger performance at creating applications from scratch.
Recent production references conflict with the announcement
People familiar with the Anthropic relationship rejected Masad’s statement and said the partnership remained in place. TechCrunch also found recent evidence across Replit’s website, blog and social posts: a late-September blog post identified Claude Fable 5 and Claude Fable 5.1 as models in Replit Agent’s production environment, while a product employee promoted the use of Opus 5.5 around the same period. The publication said it had asked Replit for clarification.
These records do not establish Replit’s current model routing. They do establish a conflict between the CEO’s description, comments from unnamed insiders and the company’s recent product references. The source does not resolve whether Anthropic models have been removed from every production path, retained in a limited role or referenced in material that no longer reflects the service.
Customers should therefore avoid treating the announcement as proof that a specific project has already moved from one model family to another. Replit’s current product documentation and behavior for the selected workflow are more relevant than a general supplier label.
Conceptual illustration of the conflict between the public statement and recent production references.
Test the app-building task, not the partnership headline
The reported decision offers a useful procurement signal: Replit says it weighs model economics, supplier cooperation and application-building performance together. It does not provide independent evidence that the resulting platform is cheaper, more reliable or better suited to an SME’s workload.
A focused evaluation can use a representative commercial task, such as modifying a lead-intake form or adding an approval step to an internal sales tool. The test should begin from the same code and requirements for each run, then examine whether the generated application follows those requirements, survives revisions and reaches a usable deployment. This tests the product outcome Masad discussed without assuming that a public benchmark represents the buyer’s own application.
The immediate decision is to treat Replit’s statement as evidence of its supplier preference, not as confirmation of a completed platform-wide migration. Before relying on model continuity or comparing costs, buyers need clarification about which models currently serve their specific Replit workflow and whether that assignment can change.
