AINEWS
Apple Opens Its AI Framework to Anthropic's Claude
Anthropic's Claude now works inside Apple's Foundation Models framework, the same on device AI interface, letting apps send requests straight to Anthropic with Apple out of the data path.
Anthropic's Claude now works inside Apple's Foundation Models framework, the same on device AI interface, letting apps send requests straight to Anthropic with Apple out of the data path.
Apple is sharing its on-device AI framework with a direct competitor. Anthropic's Claude ships as a conforming provider inside Apple's Foundation Models framework, the same programming interface that drives Apple's own on-device model. For an iOS app developer, that means a single call to the LanguageModelSession API can reach either Apple's model running locally on the device or Anthropic's Claude running in the cloud. The app picks the backend per session, and Apple sits entirely outside the request path.
The integration lives in a new Swift package called "Claude for Foundation Models," published by Anthropic as part of its Claude developer documentation. The package conforms to Apple's LanguageModel protocol, so the same LanguageModelSession surface used for Apple's model also drives Claude: respond(to:), streaming, guided generation, and tool calling all behave the same way. Developers do not write Claude-specific code.
The architectural detail is the part that gets lost in a one-line announcement. Requests go directly from the developer's app to Anthropic's Claude API. Apple is not in the request path and does not see the prompts or responses, per Anthropic's documentation for the package. Usage is billed to the developer's own Anthropic account at standard API pricing. The choice between Apple's on-device model and Claude happens at session creation, with the app making the call.
That shape matters historically. Apple has spent two decades acting as a gatekeeper on iOS: the App Store review process, the in-app purchase cut, the browser engine restrictions, and the privacy nutrition labels. Each of those moves pushed user data, payments, or rendering through Apple's servers or policies. With Foundation Models, Apple is doing the opposite. It publishes an interface, lets a competitor plug a server-side model into that interface, and steps out of the way. The framework is still Apple's, and Apple still decides what counts as a conforming provider, but inside that gate the developer can pick. That is structurally different from the App Store model and from the early-2000s browser engine lockout.
There are real limits. The package is in beta, targeting the Foundation Models server-side API introduced in the OS 27 betas. APIs and pricing can shift before general availability. The public surface is provider conformance plus configuration types, not a general-purpose Messages API client, so a developer cannot use the package for arbitrary Claude features that fall outside the Foundation Models abstraction. The on-device Apple model is the default endpoint, and Anthropic still controls the commercial relationship for the cloud path: pricing, rate limits, and the service terms. The framework gate is still Apple's, so "open" here means conformant, not unrestricted.
For a developer shipping on iOS today, the practical effect is optionality. An app that needs a stronger model for a hard reasoning task, or wants a consistent cloud fallback when the on-device model is too small, can now reach Claude without bringing in a parallel SDK or auth flow, and without making a second billing relationship visible to users. The catch is timing. This is a beta, it is OS 27 and Xcode 27 gated, and the surface area will almost certainly move before the framework ships to everyone in the fall.