Replit’s CEO says its Anthropic relationship has ended; TechCrunch reports a contrary account, leaving the status disputed.
The uncertainty makes a distinction unusually important: a company’s commercial relationship with a model supplier is not the same thing as the models described in its technical documentation. Public material can illuminate the platform’s architecture and earlier deployments without settling the terms of a present agreement.
For developers, the more useful investigation starts with what each document actually establishes. A historical engineering account, a model-selection page and an organization’s access policy answer different questions. None should automatically be treated as a live audit of every customer’s requests.
A multi-model platform, not a single switch
Replit’s model-selector documentation describes both automatic routing and manual selection. Automatic routing weighs task complexity, quality and cost, and can change models as work changes. Free Mode uses that routing rather than offering a manual model choice. Core and Pro users can select an available primary model in Power or Max Mode.
The published lists include models from both Anthropic and OpenAI. However, the documentation explicitly says availability can depend on rollout, organization settings and authorization, and that the selector visible in an account is its source of truth. A choice applies to the next message rather than proving a permanent platform-wide configuration.
This is evidence about the product as documented, not direct evidence of a supplier contract. It also cautions against assuming that every customer sees the same menu. Reading a model name on a public page cannot establish how much traffic reaches that provider, what commercial conditions apply, or whether the page reflects a very recent operational change.
Earlier deployments explain the engineering context
In its September 29 engineering post, Replit described running Anthropic’s Fable models and OpenAI’s Astra in production during August and September. That dated account is relevant background, not proof of an unchanged relationship in October. It documents a period when the company was discussing multiple model families inside its agent system.
The post also described a harness in which a main agent could select specialist subagents, reasoning effort and other execution choices. Replit evaluated configurations on named software-engineering benchmarks. Those are company-reported experiments under particular setups, not an independent finding that one vendor is superior for every application or workflow.
An August 18 announcement supplies another piece of dated context: Replit introduced Free Mode as powered by OpenAI’s GPT-5.6 Luna, alongside its other agent modes. Its claims about greater affordability and capacity were launch claims, not fresh measurements made for the current disagreement.
Together, these materials describe different layers of the system: an accessible mode, the model chosen for a task and the orchestration around that model. Treating them as interchangeable obscures what a vendor change could mean. A model comparison can inform engineering choices without revealing a partnership’s present contractual status.
Workspace policy adds another boundary
For enterprise customers, Replit’s model-control documentation describes an additional constraint. Administrators set provider and model availability at account and workspace level. Automatic routing selects from the authorized pool; it does not override those policies to use a disabled model. Different workspaces can therefore have different permitted choices.
The documentation says these controls operate at workspace level, rather than providing a different policy for each user group within one workspace. It also describes a separate setting governing fast model variants. These are practical configuration details, not evidence that any particular organization has enabled a provider or that its traffic has been inspected.
The public record therefore supports a careful account of how Replit describes model access and how it previously deployed models. It does not resolve the competing accounts of the supplier relationship. Until that point is clarified, a documented model menu should remain a documented menu, and a disputed commercial statement should remain attributed rather than being converted into a confirmed platform-wide removal.
