Code analysis shows Siri's new architecture can delegate to Claude or swap in ChatGPT

Apple's rebuilt Siri appears to have been engineered with third-party AI models as first-class citizens. According to MacRumors, a code researcher known as pdfu dug into private frameworks in iOS 27 and macOS Golden Gate and surfaced two distinct mechanisms for wiring external models such as Anthropic's Claude and OpenAI's GPT-5.6 into Siri: one that adds them as extensions, and one that appears to let them take over Siri's server-side intelligence entirely. Pdfu demonstrated both in a video shared on X.

Claude as a Siri extension

The first mechanism, which the frameworks call Model Delegation, treats a third-party model much like the ChatGPT integration Apple already ships. In the demonstration, a macOS user opens the system's Search or Ask bar, picks Claude from an Ask contextual menu, and enables the extension. Asking Siri to ask Claude to set a reminder in Apple's Reminders app then works end to end: Claude parses the natural-language request while Siri performs the action, which implies that whenever a task touches an Apple system feature, the external model hands execution back to Siri. The video also shows Claude successfully creating a CSV file inside a Siri conversation window — a task MacRumors notes Siri cannot handle on its own.

Replacing Siri's server-side brain

The second protocol, inside a component called Model Manager Services, goes considerably further. It seems designed to let Apple's own server-side Siri model be swapped out completely; the demo uses GPT-5.6, which pdfu's post refers to as GPT-5.6 Terra. In this arrangement, the replacement model receives Apple's native Siri planner prompt and tool definitions, allowing it to invoke system actions, receive the resulting personal data and produce an answer that Siri then delivers through its own interface and voice. In the demonstration, the ChatGPT-backed Siri is asked to find emails on a specific topic, summarize their contents and action points, and then message a contact through the Messages app. The resulting exchange shows up as logged activity in OpenAI's platform web interface.

A likely nudge from Brussels

MacRumors points to the European Union's Digital Markets Act as a probable influence on the design. The DMA obliges Apple to give third parties effective access to iOS hardware and software features that Apple's own services rely on, and the European Commission has stated explicitly that this principle extends to Siri. Deep model interoperability of the kind pdfu uncovered would be one way to satisfy that obligation.

What ships today

None of this is user-facing yet beyond ChatGPT. In the macOS 27 Golden Gate Release Candidate — effectively the final build of the software — the Ask menu offers only the existing ChatGPT extension, so Claude cannot actually be selected. Apple has also not opened the model delegation entitlement to outside developers, and the inference-provider path is not exposed to users. The plumbing, however, is clearly in place, and it arrives alongside iOS 27, which MacRumors reports rolls out on September 14, 2026 with what the site describes as more advanced Siri capabilities.

Why it matters

If Apple follows through, Siri stops being a fixed assistant and becomes an orchestration layer: planner prompts, tool definitions and system access supplied by Apple, with the reasoning model pluggable underneath. That would reshape the AI assistant market, since whichever model Apple features — or users choose — would gain privileged access to an enormous installed base of devices. It also reframes Apple's long-criticized lag in frontier models, because partners could supply the intelligence while Apple keeps the interface and integration. Finally, it raises real questions about trust and privacy: a replaced Siri model would be making tool calls that surface personal data, which is exactly the tension the DMA's access requirements were written to manage.