Replit chief executive Amjad Masad says the coding platform is no longer working with Anthropic, but TechCrunch’s report leaves the relationship unresolved. It records his explanation in terms of cost, partnership and performance, while also citing an unnamed person familiar with the arrangement who says the partnership remains in place. TechCrunch says it sought clarification from Replit. We cannot treat that conflicting account as a confirmed termination.
There is a separate product fact to keep in view. Replit’s model-selector documentation, as retrieved for this article, still lists Claude options alongside GPT-6 models. It says availability varies with rollout, organization settings and authorization, and directs users to their own selector for the applicable list. A documentation entry does not establish the commercial terms between two companies or guarantee access for every account.
Replit’s technical post about delegation provides historical context: it describes production observations for Fable 5 in August 2026 and Fable 5.1 and Astra in September. That supports the narrower point that its published engineering discussion included multiple models. Earlier production observations cannot settle the current contractual dispute.
Our analysis: the documentation makes the model-access question more concrete than the disputed partnership statement does. It describes Free Mode as automatically routed, while Core and Pro users can select an available primary model in Power or Max Mode. A selection applies to the next message. Those distinctions mean that “Replit uses this model” is an incomplete description of what a particular user can select.
For someone deciding whether a current project is affected, we would begin with three fields: the workspace configuration, the mode used for the task and the available primary-model choice. Recording those fields would give a support request a precise starting point. It would also distinguish an account where a choice is unavailable from one where automatic selection was used without a manual choice.
There is a second control to keep separate. Replit documents reasoning effort as a per-model setting, and says Build and Design maintain separate model selections. A comparison that records only the brand name could therefore miss a relevant configuration difference. We would record the actual controls used in each workflow before interpreting a different result as evidence that a provider had been removed.
An execution record would answer a different question again. The menu can document what was offered, but it does not show every step inside an agent run. Replit’s historical delegation post discusses a core loop handing work to subagents. A reader concerned about the provider involved in a specific task would need an appropriate execution record or a direct supplier explanation; the public menu alone cannot supply that trace.
For a practical evaluation, we would keep a compact comparison sheet for one real maintenance task: requested change, initial project state, selected controls, completed change, reviewer corrections and billed usage. The point would be to compare accepted work under recorded conditions. We have not run that comparison, and it would not resolve a commercial dispute; it would answer the narrower question of whether the user’s task still works acceptably.
Masad’s preference for OpenAI models is likewise a claim with a stated vantage point. TechCrunch reports his reference to internal and public benchmarks. We have not independently reproduced those results. A result for building an application from scratch would also leave a separate question about maintaining an existing project, so an affected user should choose the task they actually need before comparing outcomes.
The result is a clearer division of evidence. An official clarification could settle what the company means by its partnership statement. The account’s documented controls can establish what the user is offered. A reviewed task result can establish whether that particular work met its requirements. None of those records should be substituted for the others.
The useful follow-up would be a clear statement separating the commercial relationship from product availability: what, if anything, changed, when it changed and which customers or workflows are affected. Until that information is available, the conflict in the reporting should remain visible. Removing it would make the story simpler, but less faithful to the evidence.
For now, the supported conclusion is bounded: Replit’s CEO has described a change in the company’s relationship with Anthropic, TechCrunch reports a disagreement about that status, and Replit’s retrieved documentation continues to describe Claude choices with account-specific limits. These observations can coexist without establishing the complete commercial or technical picture.
If this affects your work, check the model options and settings in your own workspace, ask Replit to clarify any change, and test a representative task before planning a migration.
