Apple's revamped Siri has been engineered to work with third-party AI models at a far deeper level than the company has publicly acknowledged, according to hidden code uncovered in iOS 27 and macOS Golden Gate ahead of the software's release on Monday.
Code sleuth "pdfu" discovered private frameworks in Apple's latest operating systems that reveal two distinct mechanisms for plugging outside models into Siri, MacRumors reported. The first, called Model Delegation, allows models like Anthropic's Claude to appear as a Siri extension in exactly the same way as Apple's existing built-in ChatGPT integration. The discovery lands on the same day iOS 27 begins rolling out to users worldwide, making it one of the most consequential finds in Apple's code this year. For more context on this story, see our ongoing more AI stories.
What the Model Delegation framework does
In a video shared on X, the researcher demonstrated how a macOS user can bring up Siri's "Search or Ask" bar and select Claude through an "Ask..." contextual menu, treating the Anthropic model as a native Siri extension. Once enabled, a voice request to "Ask Claude" to set a reminder in Apple's Reminders app works end to end: Claude interprets the natural-language request, then hands the task back to Siri, which creates the reminder.
That handback mechanism appears to be the core of the design. If a request requires access to an Apple system feature, the third-party model delegates the action back to Siri rather than requesting raw access itself. In another demonstration, Claude received a request to create a CSV file inside a Siri app conversation window — something Siri itself cannot handle — and successfully returned the finished result.
Replacing Siri's server-side model entirely
The second protocol uncovered in the code goes considerably further. An inference provider inside a framework called Model Manager Services apparently allows Apple's own server-side Siri model to be replaced wholesale by another model — in the demonstration, OpenAI's GPT-5.6 Terra.
Under this arrangement, the outside model receives Apple's native Siri planner prompt and tool definitions. That means it can request system actions, receive the resulting personal data, and formulate an answer that Siri then presents using its own interface and voice. The researcher's video shows the ChatGPT-powered Siri being asked to find emails about a specific topic, summarize their contents and action points, and send a message to a contact via the Messages app — with the full exchange logged in OpenAI's developer platform interface.
In effect, Siri becomes a skin, and the intelligence underneath becomes a user's choice. Nothing in Apple's public messaging has suggested the company planned to expose this level of model interchangeability to customers.
Europe's fingerprints are all over the design
MacRumors notes that the European Union's Digital Markets Act may have helped shape Apple's approach. The DMA requires Apple to give third parties effective access to the same iOS hardware and software features available to Apple's own services, and the European Commission has specifically stated that this principle extends to Siri.
Under that pressure, engineering Siri as a model-agnostic router — rather than a walled garden tied to Apple Intelligence — gives Apple a defensible position with regulators while also future-proofing the assistant as frontier models iterate.
What is actually available today
Despite how complete the plumbing looks, almost none of it is live for users. According to MacRumors, the "Ask..." implementation demonstrated in the video is currently limited to the ChatGPT extension in the macOS 27 Golden Gate Release Candidate — effectively the final version of the software released Monday. Claude is not yet selectable, and Apple has not opened the model delegation entitlement to third-party developers.
The framework is also not front-facing: it exists in the code and can be exercised by a researcher with the right tools, but there is no user interface for choosing alternative models on an iPhone or Mac today.
That gap between capability and availability is familiar territory for Apple's AI efforts. The company spent much of the past two years rebuilding Siri around an on-device and server-side hybrid architecture, and iOS 27's headline Siri AI improvements finally began shipping with Monday's release. The discovered frameworks suggest Apple wants the architectural option to swap models without redesigning Siri again — whether to satisfy regulators, respond to competitive pressure, or both.
Why it matters for the AI industry
If Apple eventually opens model delegation and inference provision to third parties, Siri would become one of the world's largest AI distribution surfaces — a default entry point where Claude, ChatGPT, Gemini or any approved model could compete for hundreds of millions of users' everyday requests.
For Anthropic and OpenAI, that is a distribution opportunity no chatbot signup flow can match. For Google, whose Gemini model already powers features across iOS 27's rival platforms, being left out of Apple's delegation framework would be a meaningful setback. And for Apple, the design hedges its bets: if one model provider stumbles, Siri's intelligence can be swapped without users noticing anything more than a settings change.
Neither Apple, Anthropic nor OpenAI has publicly commented on the discovered frameworks. Apple did not respond to requests for comment at the time of MacRumors' report.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
