Apple's on-device AI API now has a cloud backdoor: Claude as a drop-in 'model'

One API, two very different places your prompt goes

Apple's Foundation Models framework (shipping in iOS 27 / macOS 27) is the standard Swift way to call AI on Apple platforms — the same LanguageModelSession API Apple uses for its own small on-device model. Anthropic's new ClaudeForFoundationModels package conforms Claude to that framework's model protocol, so the same respond(to:), streaming, structured output and tool calls work whether you point the session at Apple's local model or at Claude.

A reader reconstructed the shape of it from the docs: you write against the protocol once, then swap the model constructor — LanguageModelSession(model: SystemLanguageModel.default) for on-device, or a ClaudeLanguageModel(...) for the cloud tier. The crucial line, straight from the docs: "Requests go directly from your app to the Claude API; Apple is not in the request path and does not see prompts or responses." Billing goes app-developer to Anthropic, not through Apple.

"Is this actually on-device? Apple caps local models around 3B"

That framing is exactly what the thread refused to let slide.

"Apple's framework caps local models around 3B params last I looked, and Claude is way bigger than that. So either there's some hybrid setup… or this is mostly a Claude SDK in FM clothing." — a reader who benchmarks open models

He's right per the docs: this targets the framework's server-side model API. The "on-device" framing is the API surface, not the compute location.

The real story is who owns the abstraction layer

A more strategic read emerged. Apple is "right on the hot path of where developers are going to yolo slop into the ecosystem," argued one high-karma commenter, so defining a clean API that vendors like Anthropic build on — with billing flowing direct from Anthropic to the developer — is shrewd: Apple sets the standard without carrying the inference cost. Another framed it as a deliberate on-ramp: get developers used to Apple's abstraction now, so a future Apple model is a one-line swap.

Then the practitioners found the footgun. The docs warn an API key bundled in an app is extractable from the binary; the thread put it bluntly — "why would anyone ship claude keys like that", and "where does the api key reside?". The official answer is a production proxy (.proxied) so the app ships no key, but the default quick-start hands a key straight to the client.

For an iOS developer, the choice is now one argument

What changed isn't a model; it's a seam. Apple turned "which AI answers this" into a single model: parameter, and Anthropic was first to plug a frontier model into that slot. The convenience is real and so is the trap: a uniform API papers over the difference between "stays on the phone, works offline, free" and "leaves the device, costs money per call, needs a proxy." The interesting question is who fills that slot next — and whether Apple eventually fills it itself.