iOS 27 code shows Siri architected to run Claude or ChatGPT in its place

What the developer known as pdfu found in the private frameworks of iOS 27 and macOS Golden Gate isn't a user-ready feature: it's the plumbing. According to MacRumors, the code shows that Apple designed Siri's new architecture to operate with third-party AI models at a deeper level than anything the company has publicly announced. In other words, Siri was designed from the inside to be swappable.

The discovery comes on the same day as the official launch of iOS 27 (Monday, September 14, 2026), Apple's biggest software update of the year, which finally delivers the most advanced capabilities of the so-called Siri AI. The interesting detail isn't what Apple showed on stage, but what was hidden in the entitlements and protocols that haven't been opened to anyone yet.

Two protocols, two levels of depth

The code reveals two distinct mechanisms, and the difference between them matters a lot to anyone thinking about architecture.

The first is Model Delegation. It allows Claude to appear as an extension of Siri, in exactly the same way the built-in ChatGPT extension already works today. In the video demonstrated by pdfu on X, the macOS user opens the "Search or Ask" bar, selects Claude from an "Ask..." context menu, and asks Siri to "ask Claude" to create a reminder in the Reminders app. Claude interprets the request in natural language and hands the task back to Siri, which actually creates the reminder. The reading is clear: when the action requires access to an Apple system resource, the third-party model delegates back to Siri. In another example, Claude receives, inside Siri's conversation window, a request to generate a CSV file (something Siri alone can't do) and returns the result.

The second protocol goes much deeper. An inference provider inside Model Manager Services apparently allows Siri's own server-side model to be completely replaced by another model, such as GPT-5.6. In this scenario, ChatGPT receives Siri's native planner prompt and Apple's tool definitions. That means the external model can request system actions, receive the resulting personal data, and formulate a response that Siri presents using its own interface and voice.

An inference provider in Model Manager Services apparently allows Apple's own server-side Siri model to be completely replaced by another model, such as GPT-5.6.

>

-- MacRumors

In the demonstration, the user asks the ChatGPT model (inside the Siri app) to find emails about a specific topic, summarize content and action items, and then send a message to a contact via the Messages app. The response later appears logged in OpenAI's platform web interface, which shows that processing actually went through an external provider.

Why Apple would do this

The regulatory context is the piece that explains much of the engineering. The European Union's Digital Markets Act (DMA) requires Apple to give third parties effective access to the iOS hardware and software resources available to the company's own services, and the European Commission has specifically stated that this principle extends to Siri. In other words: Apple may have built swappability not out of generosity, but because it will have to prove that Siri isn't a walled garden.

An honest measuring stick is due here: none of this is enabled for the end user. The "Ask..." implementation in the macOS 27 Golden Gate Release Candidate is currently limited to the ChatGPT extension, and Claude doesn't yet appear. Apple also hasn't opened the model delegation entitlement to third parties, and the feature isn't exposed to the user. What the code proves is architectural intent, not availability.

What changes for those who build

For Brazilian developers, the takeaway is about platform design before it's about product. Based on what the code indicates, Apple treated model choice as an interchangeable abstraction layer, with a well-defined contract: the model receives the planner prompt and the tool definitions, and returns tool calls that the system executes. This is the "pluggable inference provider" pattern that many AI applications will need to adopt as swapping models (for cost, latency, or availability) stops being the exception and becomes routine.

In the Hacker News thread, this reading was exactly the one that gained traction. User akie sums up the move:

"This is a very clever platform play. If you make Siri / your mac the central access point for all AI related workflows and make models interchangeable, then Apple has a huge weapon in the AI race."

>

-- akie

Another commenter, taybin, treats the abstraction as basic engineering hygiene rather than a big novelty: according to him, "It’d be short sighted to not have a wrapper around which AI to use, even without implementation for them. Like, having an abstract parent class with a Siri implementation of it would be baseline good software engineering." The observation is useful as a counterpoint: an interchangeable model contract is, at bottom, applying polymorphism to an AI problem.

There are also those who point to a path Apple could formalize. sajithdilshan suggests the company create "a platform where app actions can be invoked with a permission model by any AI model," something, in his words, similar to MCPs for invoking app tools, published openly so that every model can be trained to use the framework. It's a reader's hypothesis, not a plan confirmed by Apple, but it aligns well with the direction the private frameworks suggest.

What remains open

The most obvious practical question is about privacy and permissions. In the GPT-5.6 example, the external model reads emails, accesses contacts, and sends messages, and the result appears logged on OpenAI's platform. What happens to the granular permission model when personal data crosses over to a third-party provider? The code shows the mechanism, not the policy.

It also remains open whether and when Apple will open the model delegation entitlement to developers beyond the already-announced partners, and whether that opening will be restricted to the European Union because of the DMA or apply globally (which would directly determine what reaches the iPhone in Brazil). For now, what we have is a snapshot of the plumbing: Siri was built to be replaceable. What's missing is for Apple to decide when, for whom, and under what rules.