1300 166 138 Tell us what you need
News / Apple

Siri could soon be running Claude or ChatGPT instead

By Monika Tantau • Published 2 months ago
Siri could soon be running Claude or ChatGPT instead

Key takeaways

• Code in the iOS 27 and macOS 27 release candidates contains two ways to run Siri on an outside model
• Model Delegation lets Claude or ChatGPT act as a Siri extension, and Model Manager Services replaces the server model
• The ChatGPT extension is present in the macOS release candidate, Claude is not, and Apple has said nothing
• The entitlement that would open this to other vendors is not available to third parties yet

Most Read

Apple shipped iOS 27, iPadOS 27 and macOS 27 this week, and someone went through the code. Inside the release candidates sit two mechanisms that let a model built by somebody else do Siri’s job: one where Claude or ChatGPT answers as an extension of Siri, and one where an outside model replaces Siri’s server side entirely.

Apple has announced none of this. The findings come from a developer reading frameworks, not from a keynote, and the part that would open the door to other vendors is still shut. That gap between what the code can do and what users can actually switch on is the whole story.

What the code actually shows

The two mechanisms are called Model Delegation and Model Manager Services, and they sit at different depths. Delegation is the shallow one: another assistant appears as a Siri extension and handles requests that come its way. Model Manager Services goes further, swapping out the model that runs on Apple’s servers behind Siri.

The deeper path matters more than it sounds. A model plugged in that way receives Apple’s own Siri planner prompt and its tool definitions, which means it is not bolted on beside Siri, it is doing the thinking Siri would have done. Claude and ChatGPT both appear in the code, and GPT 5.6 Terra turns up in the examples, according to MacRumors.

This is not the ChatGPT hand off Apple already talks about. In that one Siri decides it cannot help, asks permission, and passes the question out as a question. What the frameworks describe is the opposite arrangement: the outside model sits where Siri’s own reasoning sits, holding Apple’s prompt and Apple’s tools, and the user would have no particular reason to notice which one answered.

What is shipping and what is not

The ChatGPT extension is present in the macOS 27 release candidate. The Claude one is not. And the entitlement that would let any third party use model delegation has not been handed out, so nothing here is something a user can go and enable this afternoon.

A word on where this comes from. Release candidates are the builds Apple hands developers days before a public launch, and people read them precisely because the code arrives before the announcement does. It is a reliable way to learn what has been built. It is a poor way to learn what will ship, because plenty of what gets found is never switched on.

Apple has not commented. There is no blog post, no developer session and no timeline, which leaves two readings: plumbing laid early for a feature that arrives later, or an internal capability that was never meant to be a product. Both happen, and the code alone does not tell you which one this is.

Worth remembering that Apple already ships these releases to a very large installed base at once. Whatever Apple decides to switch on, it arrives on hundreds of millions of devices in a week rather than a year.

Why this is a bigger deal than a settings toggle

Why this is a bigger deal than a settings toggle

Siri is the one assistant on an iPhone that does not need installing. It sits behind a button, it reaches contacts, calendars, messages and apps, and most people never chose it. Everything else, including Claude and ChatGPT, lives in an app the user had to go and get.

If the model behind that button becomes a choice, the question stops being which app has the better answers and starts being which model gets the system level access. That is a different competition, and it is the one that decides what people actually use rather than what they say they prefer.

Why Australian businesses should care

Plenty of Australian teams have already picked a side, often after comparing Claude and ChatGPT on their own workloads. Those choices were made about an app, with a login, an admin console and an agreement about what happens to the data. A model sitting behind the Siri button is a different arrangement.

The practical question for anyone handling customer records under the Privacy Act is what leaves the device and who processes it. Apple splits work between on device processing and its own servers, and a delegated model adds a third party to that path. If this ever ships, the sensible first move is finding out which requests go where before deciding whether staff phones should use it.

It also matters for the teams that went the other way and kept models in house. Interest in running things locally has been climbing for exactly this reason, and our rundown of the best local LLM options covers what that actually costs. A system assistant that can route to an outside model does not change the case for local, it sharpens it.

A quieter kind of switch

Apple spent years insisting Siri was Apple’s to build. Code that lets a rival model answer through Siri’s own planner is a long way from that position, even if nobody at Apple ever says so out loud.

For businesses the useful posture is neither excitement nor alarm. Nothing has shipped, nothing is switchable, and a feature found in a framework can be pulled before anyone sees it. What has changed is that it is now worth asking your vendor, and your own policy, what happens the day the assistant on a work phone can quietly hand a question to someone else’s model.

The companies that already know which of their tools talk to which model will answer that in an afternoon. The ones that do not will find out the hard way, which is usually a support ticket about something a staff member typed into a box they assumed was Apple’s.

Scroll to Top