Buried inside the iOS 27 and macOS Golden Gate code are working Siri AI extensions that let Claude or ChatGPT take over parts of Siri’s brain — and in one demo, replace it entirely. Developer pdfu found the hooks this week and even got Claude running as a live Siri extension. Apple hasn’t confirmed a launch date, but the plumbing is already there.
- Discovery: private “Model Delegation” and “Inference Provider” APIs found in iOS 27 / macOS Golden Gate code, reported September 14, 2026.
- Who found it: developer pdfu, who demoed Claude running live as a Siri extension.
- Named partners: Claude (Anthropic) and ChatGPT (OpenAI), including a GPT-5.6 Terra reference.
- Status: unreleased and not user-facing yet — likely tied to EU Digital Markets Act requirements.
- What it would let you do: pick Claude or ChatGPT from Siri’s “Ask…” menu, or swap out Siri’s underlying model completely.
What These Siri AI Extensions Actually Do
Right now, Siri runs on Apple’s own models, end to end. The code pdfu found describes something different: a framework where an app can register itself as a Siri extension, and Siri hands off a request to it instead of answering itself.
Think of it less like a Siri replacement and more like Siri gaining a set of specialists it can call in. Ask something Apple’s model can’t handle well, and Siri could quietly route the question to Claude or ChatGPT, then hand the answer back to you inside the same conversation.
That’s a real shift for Apple, which has spent years keeping Siri’s underlying model locked to its own infrastructure (with the narrow exception of the existing ChatGPT hookup for specific queries). iOS 27’s public release already brought a smarter Siri with better personal context — this hidden layer suggests Apple was building for a lot more than that.
Model Delegation vs. Full Replacement: Two Very Different Modes
Reporting from 9to5Mac and MacRumors describes two distinct protocols buried in the code, and they behave nothing alike. Here’s how they compare.
| Aspect | Model Delegation | Inference Provider Protocol |
|---|---|---|
| What it does | Lets a third-party AI act as one Siri extension among others | Swaps out Siri’s own server-side model entirely |
| How you’d trigger it | An “Ask…” menu lets you pick a specific AI for a task | Happens system-wide, behind the scenes, every time |
| System access | Limited — hands unresolved requests back to Siri | Deep — can receive Siri’s own planner prompt and make system tool calls |
| Current status | Working demo with ChatGPT; not open to developers | Present in code only; nothing user-facing yet |
The second mode is the one worth paying attention to. If Apple ever ships it, the replacement model wouldn’t just chat with you — it would see Siri’s internal instructions and get to act on your device, the same way Apple’s own model does today.
Why Claude and ChatGPT Are the Names Showing Up
According to 9to5Mac, pdfu’s demo showed Claude working as a functional Siri extension, while ChatGPT is reportedly the only one currently reachable through the delegation path. A reference to GPT-5.6 Terra also turned up in the code, per MacRumors.
None of this is random. Apple already has a working ChatGPT partnership for Siri queries it can’t answer, so extending that same pipe to a formal extension system is a small step. Claude’s appearance is newer and more interesting — Anthropic doesn’t currently have a public Siri integration, so a working demo suggests real conversations have happened behind the scenes.
It’s worth remembering Anthropic has been shipping fast lately, including the Claude Fable 5.1 update that cut inference costs — the kind of efficiency gain that makes Claude a cheaper option for a phone-scale assistant than it would have been a year ago.
When You Might Actually Get to Use This
Nobody outside Apple has a confirmed timeline, and that matters more than the code itself. A hidden framework in a shipping OS isn’t a promise — Apple ships unused code paths all the time.
The most likely trigger is regulation, not generosity. Reports point to the EU’s Digital Markets Act, which already forces Apple to open up parts of iOS to competitors, as the probable reason this exists at all. That means EU users could plausibly see it first, if it ships at all this cycle.
There’s also the practical question of what happens when a third-party model gets Siri’s planner prompt and system access. That’s a privacy and support headache Apple will want airtight before flipping it on for anyone. Expect this to surface, if it does, as a quiet toggle deep in Settings rather than a headline feature at a keynote.
Why This Matters Even If It Never Ships
Even as unreleased code, this tells you something about where Apple thinks Siri is headed. The company that built its whole pitch on doing AI in-house, on-device, and under its own brand has apparently also built the plumbing to hand that job to someone else.
That’s a hedge, and a smart one. If Apple’s own models keep lagging the frontier, a Siri extension system means the company can swap in whoever’s winning without rebuilding the assistant from scratch. The macOS Golden Gate release shipping this same code only strengthens that read — this isn’t an iPhone-only experiment, it’s baked across Apple’s platforms at once.
Frequently Asked Questions
Can I replace Siri with ChatGPT or Claude right now?
No. The code exists in iOS 27 and macOS Golden Gate, but neither Model Delegation nor the Inference Provider Protocol is open to the public yet. Only a developer demo has shown it working.
What’s the difference between this and the existing ChatGPT integration in Siri?
Today’s ChatGPT hookup only kicks in for specific queries Siri can’t answer, and you have to approve it each time. These new Siri AI extensions describe a formal system where third-party models register as extensions or replace Siri’s model outright.
Will this feature be free?
Apple hasn’t said, and there’s no pricing information in the code. Using Claude or ChatGPT through Siri would likely still require your own account or subscription with Anthropic or OpenAI, similar to how the current ChatGPT integration works.
Why would Apple let another company’s AI run Siri?
The most credible explanation is regulatory pressure, particularly the EU’s Digital Markets Act, which requires Apple to give third parties real access to iOS. It also gives Apple a fallback if its own models fall behind competitors.
Is my data safe if a third-party model gets Siri-level access?
That’s still unknown, since Apple hasn’t published any details on how permissions or data handling would work for the Inference Provider Protocol. Given the deep system access described, expect Apple to require explicit opt-in before anything like this reaches real users.
This is the most interesting thing Apple hasn’t announced all year. A company that built its entire AI pitch on privacy and doing it themselves has quietly wired up an exit ramp for other people’s models — and that tells you Apple isn’t as confident in its own Siri roadmap as its keynotes suggest. Treat it as a feature to watch, not one to wait for; nothing here works today, and Apple could just as easily leave it dormant for another year.
Leave a Reply