Apple's on-device AI framework: helpful API or quiet lock-in?

Apple shipped its Foundation Models framework, arriving in iOS 27 and macOS 27 this fall. The GitHub repo links directly to Anthropic's ClaudeForFoundationModels, meaning Claude can be called through Apple's Swift API. That sounds convenient, but the HN thread immediately flagged the strategic angle: Apple is building a standard abstraction layer that sits between developers and every LLM, including third-party ones.

The pattern is classic Apple. Offer a clean, well-documented API that makes the easy path also the Apple-controlled path. Once enough apps are built on top of it, switching costs rise and Apple gains leverage over which models get recommended, which ones get on-device optimization, and eventually which ones get featured. The Anthropic integration is a trojan horse, or a test balloon, for a much bigger strategy.

The counterpoint in the thread: Apple has spent serious money training its own models, and this framework may simply be the staging ground before they deprecate third-party inference entirely. Either way, developers building AI features on Apple platforms need to decide now whether to go native or stay cloud-native and accept the friction.