Apple gave iPhone owners a new kind of control at WWDC26: the ability to route Siri queries to different AI models, including ChatGPT, Gemini, and Claude. The company introduced a LanguageModel protocol inside its Foundation Models framework that lets both the system and third-party apps select which backend handles a given request. For the first time, the AI powering a Siri response is no longer locked to a single provider, and developers can build apps that tap whichever model best fits a task.
Multi-model Siri changes how AI reaches your phone
The shift matters because it turns model selection into a practical choice rather than a behind-the-scenes default. Apple’s developer documentation states that builders can now “work with any language model,” listing Apple Foundation Models alongside cloud models like Claude and Gemini as options available through the same protocol. That single interface means an app could send a coding question to one model and a creative-writing prompt to another, all within the same session.
The immediate consequence for users is that the quality of a Siri answer will depend partly on which model fields the query. Claude, built by Anthropic, and Gemini, built by Google, differ in how they handle length, tone, and factual detail. Once apps begin routing requests through the LanguageModel protocol, those differences will show up in everyday use. Developers building with the framework will be able to observe how response characteristics vary across backends through their own telemetry, even if Apple has not yet published comparative benchmarks.
A hypothesis worth tracking: as the protocol reaches production apps, measurable gaps in response length and factual density between Claude-routed and Gemini-routed queries should become visible in anonymized developer logs. That data could reshape which models developers favor for specific use cases and, over time, influence which AI most iPhone owners interact with daily.
Apple’s Foundation Models framework and the LanguageModel protocol
Apple’s WWDC26 announcement positioned the update as part of its next generation of Apple Intelligence and Siri AI. The Foundation Models framework session detailed how the LanguageModel protocol works: local on-device models and server-side models both back sessions through a unified API, so developers write one integration path regardless of where inference happens.
That architecture is significant because it lowers the barrier to switching models. Instead of building separate pipelines for each provider, an app can declare a preference or let the system decide based on availability, latency, or privacy requirements. Apple’s on-device models handle lightweight tasks without sending data off the phone, while heavier queries can be routed to cloud providers through Apple’s Private Cloud Compute infrastructure.
An independent technical analysis published on arXiv examined Apple’s Private Cloud Compute design and proposed a benchmarking framework for evaluating privacy-preserving inference. The paper’s analysis and evaluation methods suggest that any cloud routing, whether to Claude, Gemini, or another service, must still satisfy Apple’s privacy standards before data leaves the device. That constraint shapes which third-party models can participate and under what conditions.
Open questions about routing, privacy, and model accountability
Several gaps in the public record leave key questions unanswered. Apple’s developer documentation and WWDC26 sessions describe the protocol’s capabilities but do not publish latency comparisons, error rates, or response-quality metrics across the available models. Without that data, developers and users are left to run their own informal tests when choosing a backend.
Equally unclear is how Apple plans to audit third-party cloud models operating inside Private Cloud Compute. The arXiv analysis offers an external evaluation framework, but Apple itself has not detailed what ongoing compliance checks Claude or Gemini must pass to remain eligible providers. If a model’s privacy behavior drifts after initial approval, the enforcement mechanism is not documented in the sources available today.
User-level opt-in mechanics also lack detail. The announcement confirms that model choice exists, but how that choice surfaces in Settings, whether users receive transparency about which model answered a given query, and whether preferences persist across app contexts are all unaddressed. Early adopters can access beta builds through Apple’s beta program, but official adoption statistics have not been released.
The practical takeaway for iPhone owners is straightforward: when the update reaches stable release, pay attention to which AI model your most-used apps select by default. Developers building on the Foundation Models framework will start making those choices on users’ behalf, and the quality of Siri-powered answers will increasingly reflect those decisions. The next thing to watch is whether Apple publishes its own performance comparisons or leaves that judgment entirely to the developer community and independent researchers.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
