Apple's Siri Can Swap in Claude or ChatGPT — What the iOS 27 Code Shows

A code researcher going by "pdfu" found something in the iOS 27 and macOS Golden Gate release-candidate builds that Apple hasn't announced: private frameworks called Model Delegation and Model Manager Services, both built to let Siri hand off requests to third-party models like Claude or GPT-5.6 — not as a fallback chatbot bolted onto the side, but as a full replacement for parts of Siri's own planning pipeline. Neither mechanism is switched on for end users, and Claude isn't enabled at all yet, but the plumbing is already built into software shipping today, September 14, 2026.

This lands the same week explainx.ai covered every AI feature Apple announced on September 9 for iOS 27's launch — Model Delegation and Model Manager Services didn't make that keynote, because they're not features yet. They're infrastructure.

TL;DR

Model Delegation: Claude as a Siri extension

The shallower of the two mechanisms works close to how the existing ChatGPT integration already behaves. In pdfu's macOS demo, a user opens the "Search or Ask" bar, picks Claude from an "Ask..." contextual menu, and asks Siri to "Ask Claude" to create a reminder. Claude parses the natural-language request, then hands the actual system action back to Siri — Reminders gets created through Apple's own app, not through anything Claude controls directly.

That handoff boundary is the whole design: Claude (or ChatGPT, currently) can reason over a request and produce an answer, including things Siri itself can't do — the demo shows Claude generating a CSV file on request, something outside Siri's native capability — but when the task needs to touch an actual Apple system feature, control routes back to Siri's own execution layer. It's the same shape as calling out to an MCP server for a capability the host application doesn't have natively, with Siri playing host.

Model Manager Services: replacing Siri's brain, not just extending it

The second mechanism is a meaningfully bigger claim. Model Manager Services exposes what pdfu calls an "Inference Providing" protocol — and under it, a third-party model doesn't sit beside Siri as an extension. It receives Apple's own Siri planner prompt and tool definitions directly, meaning the external model gets the same instructions and the same list of callable system actions Apple's own server-side model would normally use.

In the demo, GPT-5.6 running under this protocol is asked, inside the Siri app, to find emails on a topic, summarize them with action points, then message a specific contact through the Messages app. It does all three steps, and the interaction shows up logged in OpenAI's own platform dashboard — direct evidence that the request, and whatever personal data it touched, left Apple's infrastructure and reached OpenAI's. Siri still renders the final answer in its own interface and voice, but the reasoning, tool-calling, and data access underneath it belongs to the third-party model for the duration of that request.

That's a fundamentally different trust boundary than Model Delegation's handoff pattern. An extension that asks Siri to execute an action keeps Apple in the loop for anything touching the OS. A backend swap under Model Manager Services puts the third-party model directly in the data path — it sees what Apple's own model would have seen, and decides what Siri says back to you.

Why the Digital Markets Act is the actual driver here

Apple didn't build this out of platform generosity. The EU's Digital Markets Act requires Apple to give third parties effective access to iOS hardware and software features that Apple's own services already get — and the European Commission has specifically said that requirement extends to Siri. That's consistent with how Apple has handled every other DMA-forced opening: sideloading, default navigation apps, and expanded NFC access are all real in the EU and functionally inert or absent everywhere else, a pattern explainx.ai has tracked across Apple's post-Tim Cook AI strategy shifts.

The release-candidate build examined here — effectively the final version of macOS Golden Gate shipping the same day as this reporting — has the ChatGPT "Ask..." extension live, but Claude is not yet enabled, and the model-delegation entitlement isn't open to third-party developers at all. Reading the code as a signal rather than a shipped feature: Apple has built the compliance mechanism ahead of being forced to switch it on, the same negotiating posture it's taken with the European Commission on other DMA provisions where Apple has proposed narrower interoperability than regulators want and waited to see which version the Commission accepts.

What this means if you build on Claude or the OpenAI API

For developers already integrating Claude or GPT models into consumer products, this is worth watching for a specific reason: if Model Manager Services opens up, Siri becomes a distribution surface for your model choice, not just another chat window competing for attention. A user who picks Claude as their Siri backend gets Claude's reasoning applied to system-level requests — Mail, Messages, Reminders, Calendar — without installing a separate app, the same "AI as an ambient layer" outcome explainx.ai discussed in how to actually work with AI agents, applied at the OS level instead of inside a single tool.

It also raises the same tool-definition and prompt-injection questions that come up in any agentic tool-calling system: Apple's Siri planner prompt and tool definitions, once handed to an external model under Model Manager Services, become an attack surface the external provider has to defend the same way it defends any other agent given real system permissions. Nothing in pdfu's findings suggests Apple has published details on how it scopes or audits that handoff yet — which is exactly the kind of gap a formal developer entitlement program would need to close before this ships broadly.

What people are asking

Will this replace Siri's on-device model too, or only the server-side one? pdfu's findings are specific to Model Manager Services' server-side "Inference Providing" protocol — nothing in the reporting shows Apple's on-device model (the one handling requests that don't need cloud compute) being swappable the same way. That distinction matters for privacy-sensitive requests Apple currently keeps on-device or inside Private Cloud Compute; routing those to a third party would be a materially bigger change than swapping the cloud-tier model.

Does picking Claude as a backend mean Anthropic sees my Health or Messages data? Based on the demo, yes, for whatever request gets routed to it — the Model Manager Services protocol hands the third-party model the same planner prompt and tool definitions Apple's own model would use, and the resulting OpenAI dashboard log in pdfu's demo is direct evidence data left Apple's infrastructure. Neither Apple nor the labs involved have published a data-retention or scoping policy for this specific integration path yet.

Is this comparable to Amazon opening Alexa to third-party skills? Not really — Alexa Skills let third-party developers add new voice capabilities on top of Alexa's own model, while Model Manager Services replaces the model doing the reasoning and tool-calling itself. The closer comparison is a browser letting you set a different default search engine: the interface stays Apple's, but the intelligence behind it becomes a genuine choice rather than a single vendor's stack, gated on regulatory pressure rather than product philosophy.

Would this apply to older iPhones that don't support Apple Intelligence today? The code findings don't indicate a hardware change — Siri's on-device planner still has to be capable enough to trigger the handoff to Model Delegation or Model Manager Services in the first place, so devices that can't run Apple Intelligence's on-device features at all likely remain excluded regardless of which cloud model eventually answers the request.

What's still missing

1. No public entitlement. Apple hasn't opened Model Delegation or Model Manager Services to third-party developers — everything demoed relies on private frameworks reverse-engineered from shipping code, not a documented API.
2. No EU rollout confirmed. Despite being the DMA's likely target, Apple hasn't stated whether or when either mechanism activates for EU users specifically.
3. No user-facing toggle. Regular users cannot access either mechanism today, in any region — this is pre-launch infrastructure, not a beta feature behind a flag.
4. No data-handling disclosure. Apple hasn't published what happens to Siri's on-device signals (contacts, location, health data) when Model Manager Services routes a request to a third-party model's servers.

Related reading

• Every AI Feature Apple Announced on September 9: Siri AI, Audio Intelligence, and More
• Tim Cook Steps Down: John Ternus and Apple's AI Strategy
• What Is MCP (Model Context Protocol)? A Complete Guide
• What Are Agent Skills? A Complete Guide
• How to Actually Work With AI Agents: A Practical Communication Guide
• Mustafa Suleyman's "Humanist AI" Code of Conduct: The 10 Rules for MAI Models

This post reflects code findings reported by researcher "pdfu" as of September 14, 2026, from iOS 27 and macOS Golden Gate release-candidate builds. Neither Model Delegation nor Model Manager Services was enabled for end users at time of writing — check Apple's official developer documentation for any formal entitlement program that may follow.