Siri may be preparing for a significantly more modular AI future. References found in private frameworks for iOS 27 and macOS Golden Gate indicate that Apple’s assistant architecture has been designed to work with outside models at more than one level. One path resembles the existing ChatGPT integration, where a user can intentionally pass a request to another assistant. The other appears potentially broader: an external model could, in theory, become the intelligence handling Siri’s server-side planning while Siri remains the familiar interface that carries out and presents the result.
That distinction matters. Adding a chatbot to an assistant is one thing; allowing a different model to reason over a request, select system tools, receive the results, and send an answer back through Siri’s voice and interface is a much larger technical proposition. The code does not establish that these possibilities are publicly launching now, nor does it mean every model named in the frameworks will become a selectable option. It does, however, point to Apple having laid substantial groundwork for AI model interoperability.
A possible “Ask…” route for Siri extensions
One mechanism is described as Model Delegation. It appears to provide a framework through which a third-party AI can live alongside Siri as an extension, similar in concept to the ChatGPT extension already built into Apple’s software.
A demonstration based on the discovered code shows Claude appearing in a macOS “Search or Ask” experience. From an “Ask…” contextual menu, the user selects Claude as the model to handle a particular request. Rather than making Siri disappear, this setup makes the outside model feel like a specialized teammate brought into the conversation when the job calls for it.
In the example, a user directs Siri to ask Claude to create a reminder in Apple’s Reminders app. Claude interprets the natural-language instruction, while Siri ultimately creates the reminder through the operating system. That handoff is arguably the most interesting part. It suggests a split of responsibilities in which an external model may understand the intent and shape the request, but Siri retains responsibility for access to protected Apple features and for actually performing device-level actions.
Put less politely, it is the AI equivalent of letting a guest write the quest walkthrough while Siri still holds the controller. Claude may know what should happen; Siri is the one permitted to press the buttons inside Apple’s ecosystem.
A separate demonstration shows Claude receiving a request in a Siri app conversation to create a CSV file. Siri itself is not able to produce that kind of result, while Claude returns the requested file. This offers a useful illustration of why delegation could be valuable beyond novelty. Different models have different strengths, and an assistant that can route a task to a capable provider could cover gaps without pretending one system excels at everything.
The deeper possibility: replacing Siri’s inference provider
The more consequential framework reference is tied to something called Model Manager Services. Within it, an “inference provider” appears capable of substituting another model for Apple’s own server-side Siri model. One example referenced in the material uses GPT-5.6.
If this works as the framework language suggests, the outside model would not merely answer a question in a separate chat pane. It could receive Siri’s planning prompt and its tool definitions. Those tools are the connective tissue that lets an assistant attempt system actions: asking for information from Apple services, requesting an operation, examining the result, then determining what to say next.
In that arrangement, the external provider could formulate the plan and request the action. Siri would continue to serve as the shell around it, presenting the completed response in Apple’s own interface and potentially using Siri’s voice. It is a model that keeps the visible Apple experience intact while making the underlying reasoning engine more interchangeable.
There are several caveats worth keeping in view. Private frameworks can reveal intended directions, experiments, or infrastructure that is not ready for release. A named model can be a development example rather than a product commitment. And any system that lets a third-party model interact with tools capable of retrieving personal information or invoking system functions raises difficult questions about consent, permissions, data handling, reliability, and the limits Apple would enforce.
So this should not be read as a promise that iPhone, iPad, or Mac owners will soon open Settings and replace Siri with the chatbot of their choice. The evidence is much narrower: Apple appears to be engineering software plumbing that could support an arrangement far more flexible than a single built-in assistant.
Why the handoff model could matter
The value of this design is not simply that it creates a menu full of AI brand names. A well-structured delegation system could let users get the benefit of an outside model’s particular skills while preserving the conventions that make operating-system assistants useful: consistent permissions, system-aware actions, a common voice interface, and a controlled way to move between apps.
- Complex interpretation: A provider could parse a detailed request that depends on strong language understanding or structured reasoning.
- System execution: Siri could be the broker that performs actions in Reminders and other Apple features where permissions and device integration matter.
- Specialized output: An outside model could generate content or formats Siri does not natively handle, such as the demonstrated CSV-file request.
- A consistent front end: Users would not necessarily need to abandon Siri’s interaction model just to use a different AI capability.
That is also why the design is more interesting than a conventional “send this prompt to ChatGPT” button. A simple handoff sends text elsewhere and returns text. The broader inference-provider concept appears to contemplate a loop in which the model can call defined tools, receive the resulting information, and continue working toward an answer. The operating system would remain central to the action layer even if another model supplied some or all of the reasoning.
For Apple, that could be a way to balance user choice with platform control. For users, it could make Siri less of a single, fixed product and more of an orchestrator. Whether that turns into a practical, understandable experience depends on the details that are not visible here: how providers are selected, which tasks may be delegated, how requests are disclosed, what information leaves Apple’s systems, and how errors are explained.
What is actually available now
At present, the “Ask…” implementation in the macOS 27 Golden Gate Release Candidate is limited to the existing ChatGPT extension. Claude is not currently available as a user-facing choice. Apple also has not opened the Model Delegation entitlement to third parties, and the capability is not presented as a public feature that people can simply enable.
The Golden Gate Release Candidate is effectively the final pre-release build ahead of the software’s scheduled release on September 14, 2026. That status makes the references notable, but it does not erase the difference between internal architecture and consumer availability. A release candidate can contain foundations for features intended for later use, including components that are inactive, restricted, or still being evaluated.
Apple’s wider iOS 27 work has already put Siri-related capabilities in focus. For another example of how apps are positioning themselves around those changes, see Things 3.24’s Siri AI and Spotlight additions on iOS 27. The newly surfaced framework details suggest Apple may be considering not just smarter built-in actions, but a broader system for deciding which intelligence helps complete them.
Regulatory pressure may be part of the backdrop
The European Union’s Digital Markets Act may be relevant to why this architecture appears so open-ended. The law requires Apple to provide third parties with effective access to iOS hardware and software capabilities that are available to Apple’s own services. European Commission guidance has specifically indicated that this principle extends to Siri.
That does not prove the framework was built because of the DMA, and the code alone cannot establish Apple’s policy plans across regions. Still, a design where qualified third-party models can interact with Siri through defined protocols would fit the general direction of rules intended to prevent a platform owner from reserving key assistant capabilities solely for itself.
For now, the practical takeaway is cautious but compelling. ChatGPT remains the visible extension in macOS 27 Golden Gate’s current implementation. Claude’s appearance in a demonstration and GPT-5.6’s apparent role as an inference-provider example show what the framework may support rather than what users can choose today. Yet the underlying message is clear: Siri’s next stage may not be about one assistant winning every task. It may be about Siri becoming the referee, tool runner, and familiar face for a roster of AI models waiting just off the bench.
