Siri is quietly becoming a router: Claude and ChatGPT found in code

5 min read 1 source clear_take
├── "Apple is building a 'search engine picker' for AI — owning the surface while making model labs compete for the slot"
│  ├── MacRumors (MacRumors) → read

MacRumors surfaced code paths showing a `modelProvider` field, per-provider capability flags, and a UI affordance resembling Safari's search-engine picker. The reporting frames this as provider abstraction plumbing that goes beyond the 2024 single-vendor ChatGPT hand-off, suggesting Apple is architecting a pluggable slot rather than another bilateral deal.

│  └── top10.dev editorial (top10.dev) → read below

Draws the Safari-2003 analogy: Apple didn't build a search engine, it built the browser and made Google the default with a dropdown for alternatives, which both placated regulators and turned Apple into the toll collector on a market it didn't have to compete in. The editorial argues the assistant on your phone is about to be the same shape — Apple owns intents, permissions, and billing while model vendors fight for the underlying slot.

└── "Code shipping before strategy means the wiring diagram itself is the announcement"
  ├── top10.dev editorial (top10.dev) → read below

Notes that Apple ships code before it ships strategy 'roughly never,' so the existence of a multi-provider routing layer being actively tested with at least two vendors is itself the substantive signal. A single-provider deal is a contract; a pluggable provider slot is a market — and Apple is clearly building the latter regardless of which names appear in the final launch dropdown.

  └── @tosh (Hacker News, 200 pts) → view

By submitting the MacRumors piece to HN where it hit 200 points and 116 comments, tosh signal-boosted the framing that leaked build-tree strings constitute meaningful product intelligence. The submission's traction reflects a shared read that this plumbing-level detail matters more than typical rumor-mill fodder.

What happened

MacRumors surfaced strings and code paths in a recent Apple internal build that show Siri is being wired to route requests to third-party model providers — specifically Anthropic's Claude and OpenAI's ChatGPT — alongside Apple's own on-device and Private Cloud Compute models. The plumbing goes further than the ChatGPT hand-off Apple shipped with Apple Intelligence in 2024. This is a provider abstraction: a `modelProvider` field, per-provider capability flags, and a UI affordance that looks a lot like the search-engine picker in Safari.

Nothing about this is a product announcement — it's a wiring diagram leaking out of a build tree — but the wiring diagram is the story. A single-provider deal (Apple ↔ OpenAI) is a contract. A pluggable provider slot is a market. Those are very different postures, and Apple ships code before it ships strategy roughly never.

The code doesn't say which providers will be user-selectable at launch, whether Anthropic has signed anything, or how billing flows. It also doesn't say whether the on-device Apple model gets first refusal (likely yes, given latency and cost) with cloud providers as fallback for hard queries. What it does say is that the routing layer exists, has slots for multiple names, and is being tested with at least two of them.

Why it matters

The useful frame here is Safari, circa 2003. Apple didn't build a search engine. It built a browser, made Google the default, took a cut, and left a dropdown so users could switch to Yahoo or Bing or DuckDuckGo. That dropdown did two things at once: it satisfied regulators nervous about default-lock-in, and it turned Apple into the toll collector on a market it didn't have to compete in. The assistant on your phone is about to be the same shape: Apple owns the surface, the intents, the permissions, and the billing relationship; the model vendors compete for the slot underneath.

That framing changes what "winning" looks like for the model labs. If you're Anthropic or OpenAI, being the default Siri backend in the US is worth an eye-watering amount — comparable to being Safari's default search engine, which Google has paid roughly $20B/year to keep. If you're not the default but you're in the picker, you're fighting for the sliver of users who change defaults, which historically is single-digit percentages. And if you're not in the picker at all — Gemini, Llama, Mistral, DeepSeek — you are, functionally, off the iPhone as an assistant.

The second thing this reframes is Apple's AI strategy narrative. For two years the take has been "Apple is behind on AI." That take assumed Apple needed to ship a frontier model to matter. If the routing layer works, Apple doesn't. Apple's moat was never the model; it was the 2.2 billion active devices and the intents API that decides what an assistant is actually allowed to do on them. A model that can't book your calendar, read your messages, or trigger a Shortcut is a chatbot. A model that can is an assistant. Apple controls which one you get.

Third — and this is the part the coverage keeps missing — a provider-swappable Siri is a hedge against exactly the kind of provider blow-up that a single-vendor deal makes catastrophic. If OpenAI has an outage, a safety incident, or a pricing renegotiation, the ChatGPT-only Siri path becomes a liability the moment it ships. A multi-provider Siri makes any single vendor replaceable at the config level. That's not a friendly posture toward partners; it's the posture of a platform that has learned from twenty years of watching apps get sherlocked and search deals get renegotiated.

The fourth signal worth noting: putting Claude in the code alongside ChatGPT is Apple explicitly declining to give OpenAI the exclusivity OpenAI would obviously prefer. Sam Altman's public posture toward Apple has been warm; his private posture toward being commoditized behind someone else's UI is presumably less so. Anthropic being in the picker means Apple negotiated with two labs against each other, which is the leverage position Apple always wants and rarely admits to wanting.

What this means for your stack

If you build for iOS, the practical takeaway is: stop targeting a specific model and start targeting App Intents cleanly. The assistant that reaches your app is going to be provider-agnostic from your app's point of view. Your Intents need to work whether the reasoning is happening on-device, in Private Cloud Compute, at Anthropic, or at OpenAI. That means:

- Keep intent schemas tight. Don't rely on the model "figuring out" ambiguous parameters — different backends will figure differently, and your users will file bugs against you, not against Apple. - Assume tool-call quality varies by backend. Aaron Patterson's rant earlier this year about malformed agent requests wasn't wrong; it was early. If Siri round-robins across providers, your endpoints will see the union of every provider's malformed-call failure mode. Validate hard, fail loud, and log the `User-Agent` equivalent so you can attribute breakage. - Design for on-device first. The Apple Intelligence model is small, fast, and free at the point of use. If your intent can be satisfied by the on-device model, it will be — cloud hand-off is a fallback, not the default. Optimize your intent surface to be answerable with short context.

If you build agents or model-facing infra outside the Apple ecosystem, the read is different: the assistant layer is consolidating into two or three surfaces (Apple, Google, whatever Meta ships), and each surface is going to have a provider picker rather than a single model. That's good news for model competition and bad news for anyone hoping to sell "agentic OS" as a startup category. The OS vendors are building the router. You're building either a model that fits in the picker or a workflow that lives above the picker in a native app.

For infra teams at the model labs: the ranking function that decides which provider Siri picks is the new benchmark. It won't be MMLU. It'll be some blend of latency, cost, refusal rate on Apple's safety taxonomy, and tool-call reliability against Apple's intent schema. Start measuring against those axes now, because the eval that matters just moved.

Looking ahead

The piece to watch is not whether Claude or ChatGPT ships as a Siri backend — the code says both are in play — but which one Apple picks as the *default*. Default placement is where the money is and where the user-behavior data goes. If Apple picks its own on-device model as default with third parties as opt-in, that's a Safari-in-2003 move: Apple keeps the relationship, vendors compete for the fallback. If Apple picks a third party as default, that's a search-deal move, and it will be worth billions to whoever wins it. Either way, the era of "which AI assistant do you use" is ending on iOS. The assistant is Siri. The question is who's answering behind it — and Apple just wrote the code that makes that question a config setting.

Hacker News 207 pts 139 comments

Apple's Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows

→ read on Hacker News
Kuyawa · Hacker News

Siri was the first widely available AI bot but they forgot to open it to developers. The ecosystem would have grown a million fold in all these yearsNever underestimate the power of a bored programmer Siri turn on the lights Siri code a Doom clone in Swift Siri design a landing page for a Consulting

taybin · Hacker News

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.

TheDong · Hacker News

And you can also sideload apps (only in the EU), change you default navigation app from apple maps (only in the EU), and use more features of the NFC reader (only in the EU).The code for all of those things is there, but not enabled if for most iOS users. There's a non-zero chance this code is

cdrnsf · Hacker News

My favorite Siri behavior is telling it, on my phone, to turn on a focus state only for it to transfer the request to a nearby HomePod that tells me it can't help with that. I hope the new changes to it are an improvement, but the bar is currently buried in the dirt.

akie · Hacker News

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.

// share this

// get daily digest

Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.