What iOS 27’s Reported Siri Code Reveals About Claude, GPT-5.6, and Model Delegation

The key finding: Siri may be built as an AI coordinator

Code researchers and subsequent reports describe two private mechanisms in iOS 27 and the macOS Golden Gate release candidate that could make Siri less dependent on a single Apple-controlled reasoning model. In this design, Siri is not merely the model answering a question. It can be the layer that receives a request, supplies tools, carries out authorized device actions, and presents the result—even when another provider performs the language interpretation or planning. 34

That distinction matters. It would make a third-party model an integrated participant in an assistant workflow rather than a standalone chatbot opened alongside Siri.

Model Delegation: an outside model interprets, Siri acts

The first reported mechanism is Model Delegation, a private App Intents-based path for Siri extensions. In a demonstration cited by MacRumors, Claude appeared in Siri’s “Ask…” menu through a flow resembling the existing ChatGPT handoff. The user asked Claude to create a reminder; Claude interpreted the natural-language request and Siri created the reminder in Apple’s Reminders app. 316

In practical terms, the division of labor appears to be:

1. A user invokes Siri and selects or addresses an external model.
2. The external model interprets the request and determines what needs to happen.
3. When the request requires an OS or app action, it hands an action back through Siri’s system tool path.
4. Siri executes the action and returns the result in its normal experience. 316

If Apple were to ship this mechanism publicly, it could let people benefit from a provider’s particular writing, reasoning, or conversational strengths without requiring that provider to replace Siri’s voice interface or directly own the system-action layer.

The deeper path: replacing the server-side reasoning model

A second reported mechanism is more consequential. The researcher’s demonstration describes an Inference Providing protocol in Model Manager Services, implemented through a private ExtensionKit extension point. In that example, an external model labeled GPT-5.6 Terra replaced Siri AI’s server-side model. 15

According to the demonstration, the provider receives Apple’s native Siri planner prompt and tool definitions, can request tool calls, receives tool-call results, and sends a final response back through Siri’s UI and voice, including interactive elements. 15 Reports characterise this as replacing a reasoning or planning component—not replacing the entire Siri product. Siri can still provide the interaction surface and the mechanism for executing tools. 415

This is a more integrated architecture than a simple “ask a chatbot” fallback. It potentially separates the assistant into components:

• Siri: invocation, interface, voice, tool runtime, and action execution.
• External inference provider: request interpretation, planning, and response generation.
• Apple-controlled tools and permissions: the path through which system actions and results are handled.

Personal context is an implication, not a confirmed public entitlement

The reports are notable because a planner that can receive tool definitions and tool results could potentially operate in workflows involving Siri’s broader context and actions. But the private-code discovery does not establish that third-party providers currently receive unrestricted access to Mail, Messages, Photos, files, or other sensitive user data.

Access to personal information and consequential actions would depend on the permissions, data boundaries, and policy Apple chooses if it ever opens the system. The demonstration shows possible plumbing, not a public privacy model or an announced third-party data-access policy. 15

What is actually available today

For users and ordinary app developers, these findings do not create a new setting to choose a Siri “brain.” The reported Model Delegation and inference-provider hooks are private, and Apple has not opened the relevant entitlement or announced them as a public developer capability. Reports on the macOS 27 Golden Gate release candidate say ChatGPT remains the only publicly available Siri handoff. 1018

That means there is currently no supported path to install Claude, GPT-5.6, Gemini, or another service as a Siri extension or server-side replacement. A video demonstration of private interfaces is evidence of an experimental or internal capability; it is not evidence of consumer availability.

How this differs from Apple’s public WWDC developer messaging

Apple’s public WWDC 2026 developer material does describe expanded App Intents capabilities. Developers can connect apps to Siri AI features such as personal-context understanding, app actions, and onscreen awareness. Apple also said developers can use models of their choice, including Claude and Gemini, through its language-model protocol. 20

Those announcements should not be conflated with a public framework that lets users select Claude, ChatGPT, Gemini, or Grok as Siri’s underlying model. Available reporting specifically says that the proposed broad third-party Siri Extensions framework was not announced and that ChatGPT remained the public Siri handoff option. 18

In other words, Apple’s public APIs describe how apps can participate in Siri-enabled experiences and how developers can use outside models in their own apps. The private frameworks reportedly describe a deeper possibility: third-party models participating directly in Siri’s own planning pipeline.

The possible DMA connection

The EU Digital Markets Act is relevant context, but it is not proven to be the source of these private frameworks. Article 6(7) requires designated gatekeepers to provide third parties access to the operating system’s hardware and software features on terms comparable to those available to the gatekeeper’s own services. 31

A model-delegation architecture could be compatible with the general idea of interoperability: an outside assistant could reason over a request while the platform supplies controlled access to system capabilities. But neither the reported code nor Apple’s public statements establish that Model Delegation or Model Manager Services was built specifically to satisfy the DMA.

What to watch next

The strongest conclusion is architectural, not commercial: the reported code suggests Apple may be designing Siri as an orchestration layer that can separate who reasons from who executes actions and owns the user experience. 34

For this to become meaningful to users, Apple would still need to publicly document the framework, open entitlements, define consent and data-sharing rules, establish review and security requirements, and make provider choices available in the product. Until then, Claude-in-Siri and replacement Siri models remain reported capabilities in private code—not features people can enable today.