Regulators, Not Competitors, Just Made Siri Model-Agnostic

Apple shipped iOS 27 on September 14 with Siri running on Google's Gemini by default. [2][3] Buried in the same release: code that lets Siri's entire backend be swapped for Claude, ChatGPT, or a model Apple didn't build at all. [4]

Nobody at Apple demoed that second part. A code researcher found it. And the reason it exists traces back to a fight Apple picked with Brussels three months earlier — a fight Apple lost on its own terms. [1] If you've ever bet a client build on "the AI inside the platform," this is the clearest evidence yet that the AI inside the platform is becoming a setting, not a commitment.

What shipped September 14

iOS 27 and macOS Golden Gate rolled out worldwide on September 14. [3][4] The headline feature is Siri AI — Apple's own description calls it "a profoundly more capable and conversational assistant" that reads your messages for context, sees what's on your screen, and drafts writing from a description. [2] TechCrunch's review confirms the assistant is now built on Google's Gemini models, replacing the ChatGPT integration Siri leaned on before. [3] The reviewer, who'd given up on Siri entirely, said he's using it again — genuinely.

That's a big swap to make quietly. Apple picked a single outside model to run its most-used assistant, and most users won't ever know that changed.

The code Apple didn't demo

The bigger story showed up somewhere Apple didn't put it: in the compiled code itself. Code researcher "pdfu" found two separate mechanisms hidden in iOS 27 and macOS Golden Gate for handing Siri off to an outside model. [4]

The first is the Model Delegation API. It puts a third-party model into Siri's "Ask…" menu as an extension — ask a question, Siri routes it to the extension, the extension answers, Siri executes whatever system action comes back. ChatGPT is live there right now, in the macOS Golden Gate release candidate. Claude shows up in the same code path but isn't switched on for users yet — the entitlement that unlocks it hasn't shipped. [4]

The second is bigger: Model Manager Services, running an Inference Provider Protocol that replaces Apple's own server-side Siri model entirely. In the code researcher's example, a model called GPT-5.6 Terra receives Apple's own Siri planner prompt and tool definitions — the same instructions Apple's model gets — and calls Siri's tools directly, requesting system actions and returning results through the normal Siri interface. Apple's model doesn't touch the request. Apple just renders the voice and the UI. [4]

Neither mechanism is open to outside developers yet. But the plumbing is built, tested enough to demo, and sitting in a shipping OS.

flowchart TD
  U[You ask Siri something] --> M{Which path handles it?}
  M -->|"Ask... menu (Model Delegation API)"| D["Extension answers.<br/>ChatGPT: live now.<br/>Claude: in the code, not yet open."]
  D --> S1[Siri executes the system action]
  M -->|"Full backend swap (Inference Provider Protocol)"| P["Outside model gets Siri's own<br/>planner prompt + tool definitions"]
  P --> S2[Model calls Siri's tools directly —<br/>Apple just renders voice + UI]
  S1 --> R[One Siri interface either way]
  S2 --> R

Why this exists: the fight Apple picked with Brussels

None of this is Apple volunteering openness. It's Apple's answer to the EU's Digital Markets Act.

On June 8, Apple announced Siri AI wouldn't launch on iPhone, iPad, or Apple Watch in the EU alongside iOS 27, citing the DMA. Craig Federighi, Apple's software engineering chief, called it plainly: "We're deeply disappointed that our EU users won't have Siri AI on iPhone or iPad when we share our new software releases later this year." [1]

Apple's own account of why: the DMA, as Apple reads it, would require giving any virtual assistant near-unlimited device access — reading messages, sending them, making purchases, moving files — without the safeguards Apple wanted in place first. Apple says it built a "Trusted System Agent" as a proposed middle ground, letting rival assistants reach the same system capabilities safely, and offered an 18-month rollout to get there. The EU didn't accept it. [1]

That's the dispute that produced the code found in September. Model Delegation and the Inference Provider Protocol read like the technical answer to "how do we let a competing assistant do what Siri does" — built for a regulatory fight, then generalized into something that ships everywhere, including the market where Apple didn't have to build it at all.

What ships where

The regional split makes the regulatory story visible in the product itself:

Mac and Vision Pro aren't designated "gatekeeper" platforms under the DMA, so Apple shipped Siri AI there without a fight. [1] iPhone, iPad, and Watch are — and that's exactly where Apple is still withholding the feature rather than open it on the EU's terms.

Story: MacRumors — Apple's Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows.

What this means if you build on someone else's assistant layer

I tell clients not to wire critical logic into a single vendor's proprietary layer without an exit plan. Systems built to connect your own tools instead of renting someone else's black box don't have this problem, because there's no vendor deciding for you which model runs your logic tomorrow.

Apple just did the large-scale version of that lesson in public. The company that has the most reason on earth to keep its assistant closed — Siri is core brand, core UX, core reason people stay on iPhone — built the machinery to swap its own model out. Not because a competitor pressured it into a better product. Because a regulator required it as the price of shipping features at all. The Hub's stack-reduction work cut its automation bill 60% by owning the integration layer instead of renting one — the same principle, at a very different scale: the company that owns the routing decision keeps the upper hand, and the company forced to give that up loses control over its own defaults whether it wanted to or not.

Apple didn't choose to make Siri's brain swappable. A regulator wrote that requirement, and now it's in the code.

The operator takeaway isn't "go build on Siri Extensions" — that door isn't even open yet. It's narrower: the same routing-layer risk I flagged when SpaceX bought Cursor and Stripe went after OpenRouter applies here too. Whoever controls the layer between the user and the model controls the defaults, the pricing, and eventually which vendor gets favored. That was true when it was two AI-infrastructure startups getting acquired. It's just as true when it's Apple, forced open by a government instead of a buyer.

If your product roadmap assumes any platform's default assistant stays fixed to one vendor, this is the week that assumption got weaker evidence than it's ever had.

Sources

[1] Apple — Due to DMA, Siri AI delayed in EU for iOS 27 and iPadOS 27 — apple.com

[2] Apple — Siri AI: a profoundly more capable and personal assistant is here — apple.com

[3] TechCrunch — With iOS 27, I'm actually using Siri again — techcrunch.com

[4] MacRumors — Apple's Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows — macrumors.com

The short version

• iOS 27 and macOS Golden Gate shipped September 14 with Siri now running on Google's Gemini by default. [3]
• Hidden in the same code: a Model Delegation API (ChatGPT live now, Claude visible but not yet enabled) and an Inference Provider Protocol that can replace Siri's backend entirely with an outside model. [4]
• Neither mechanism is open to third-party developers yet — it's demoed, not shipped as a feature.
• This plumbing traces back to Apple's June 8 fight with the EU over the Digital Markets Act, which Apple lost on its own proposed terms. [1]
• Siri AI ships on iPhone, iPad, and Watch everywhere except the EU — where only Mac and Vision Pro, the non-gatekeeper platforms, have it.
• If your business logic depends on a platform's default AI staying fixed to one vendor, that bet just got visibly weaker.