MacRumors published a report on Monday that climbed to 83 points and 25 comments on Hacker News — solid traction for a pure reverse-engineering post. The core finding: iOS code contains interface scaffolding that could swap Siri's backend AI model for third-party options like Claude or ChatGPT.
Published September 14, HN item ID 49695409. I went through the original post and the comment thread. Here are the points worth unpacking.
What the Code Actually Shows
MacRumors' reverse engineering focused on Apple's private frameworks, specifically the Backends and ModelAdapter classes along the Apple Intelligence call chain. Key strings referenced in the report include:
ChatGPTSiriBacked— clearly pointing to OpenAIClaudeSiriBacked— the typo is from the original report, likely truncated strings from the binary; almost certainlyBackend- A generic
RemoteModelProviderprotocol with an internal enum supportingapple,openai,anthropic, andcustom
This goes beyond "ChatGPT integration" — Apple Intelligence already routes certain queries to ChatGPT with user consent. What was found here sits much deeper in the stack. The RemoteModelProvider abstraction suggests Siri could route user-intent parsing directly to non-Apple models as a primary inference backend, not merely as a fallback.
Some details worth flagging:
- The strings don't go through
NSLocalizedString, meaning they're backend routing logic, not user-facing UI copy. - The classes live in private frameworks, unreachable by normal apps — only system services can hit them.
- The
customenum value is particularly interesting — it implies Apple is considering enterprise users or specific scenarios plugging in self-hosted models.
The report's own wording: "The code suggests Apple is preparing to allow users to choose which AI model powers Siri, with third-party options beyond the current ChatGPT integration." The word "preparing" matters — nothing has shipped; the hooks are just in the code.
Why This Isn't Just "Third-Party Integration"
Two very different scenarios are on the table:
A. Fallback mode: Siri's primary path stays on Apple's own model. Only when it hits something the in-house model can't handle does it fall back to ChatGPT or Claude. That mirrors the current ChatGPT integration in iOS 18 — just deeper and more automatic.
B. Model-layer abstraction: Siri's inference backend is designed to be pluggable. Users pick a default model in Settings, or different tasks route to different models — simple commands on-device, complex reasoning offloaded to a cloud Claude.
The RemoteModelProvider protocol plus the custom enum lean toward B. But MacRumors stayed cautious — "suggests," no hard conclusion. My read is that Apple is testing both paths in parallel, and the final product shape depends on how negotiations with Anthropic and OpenAI shake out.
The distinction matters. If it's A, Apple Intelligence's strategy hasn't changed — own models first, third-party as patch. If it's B, Apple is effectively conceding that its general-purpose models can't match Claude and GPT in the near term. That's a significant pivot for the whole Apple Intelligence roadmap.
What the HN Thread Was Buzzing About
Three representative comments:
@bstar (41 points): "If Shortcuts can invoke different models per action, this changes everything for automation builders." — Pointing at Shortcuts and App Intents integration. If each Shortcut action could pick its own model, automation quality jumps: small model for "summarize email" to save cost, Claude Opus 4 for "analyze earnings report" for quality.
@kalleboo: "The string ANECompiledModel appears nearby in the same binary. They're not just routing to cloud — looks like hybrid inference is on the table." — ANE is the Apple Neural Engine. The implication: Apple is weighing a hybrid on-device + cloud inference architecture, with simple tasks handled locally and complex ones escalated to Claude or GPT.
@throwaway (counterpoint): "Remember when Apple blocked Facebook's internal apps? They will absolutely gate this behind a curated whitelist. You won't get to plug in your own Llama." — A reminder not to get carried away. Apple's version of "open" has always been controlled. Third-party models would almost certainly pass through Apple review — not a real open ecosystem.
The mood was excited but measured. Developers smelled an opening in Apple Intelligence's roadmap, but most are waiting for an official WWDC-style confirmation before celebrating.
What This Actually Means for Anthropic and OpenAI
News traction is one thing; real call volume is another. Estimating from three variables:
iOS install base: Over 1.3 billion active iPhones globally. But Siri's daily active rate isn't officially disclosed — industry estimates put it at 30–50%, so 400–600 million daily users. Of those, the share willing to switch the default model to Claude might land around 5–10%. Realistic call volume: a few hundred million per month.
User willingness to pay: The real bottleneck. Siri is currently free. Getting users to pay extra for a Claude backend won't convert well. Apple will probably run a "free with quota, then subscription" model, or subsidize the basics to seed adoption. What Anthropic actually gets isn't cash — it's brand exposure and user-data collection opportunities.
Apple Intelligence's privacy posture: Apple has pitched "Private Cloud Compute" as a core promise — user data isn't used for training. If a Claude backend has to slot into Apple's privacy architecture, Anthropic would need to compromise on data handling that conflicts with how it sells API access today.
Net assessment: for Anthropic and OpenAI, iOS Siri integration is strategically significant but not a near-term revenue play. The biggest beneficiary may be Apple itself — by slotting in "model supplier" as a role, it pressures its own small-model team to ship better results, or watch users vote with their fingers.
Editor's Take
What sets this apart from past "big-model integration" stories is a shift in how Apple positions itself.
Microsoft Copilot runs on GPT — Microsoft is an OpenAI shareholder, so it's essentially one company using technology from another company it owns. Google Gemini defaults to Google's own models — fully closed-loop, nothing to question. Amazon slots Claude into Bedrock — a cloud vendor running a model marketplace, a neutral-platform play.
Apple is different. Apple isn't a major shareholder of either Anthropic or OpenAI; it has no deep binding tie to either. If users genuinely get to pick Claude or ChatGPT as their Siri default, Apple is acting as a model-neutral platform — and that collides with two decades of "hardware-software vertical integration" identity.
Two readings I can construct:
Reading A: Apple is paving the way for on-device / cloud tiering. The ANE + remote-model hybrid architecture @kalleboo mentioned in the HN thread is the real target. Local small models handle routine queries; cloud large models handle complex ones — and that cloud model is Apple Intelligence's own. The "swappable" interface exists only to bridge the gap until Apple's own models catch up, at which point the interface narrows. This fits Apple's historical behavior pattern best.
Reading B: Apple is being forced open. If Apple Intelligence keeps lagging behind Claude 4 / GPT-5, staying closed just drives user churn. iOS loyalty is high but not infinite. Admitting reality and handing users the choice might be the better way to preserve ecosystem stickiness.
I'm closer to A, but the uncertainty is the timeline. If Apple's own models don't close the gap within 2–3 years, Reading B starts to dominate.
One thing I couldn't verify: what the custom enum value actually means. Enterprise users plugging in self-hosted models? Internal Apple testing? A future "developer mode" escape hatch? MacRumors didn't elaborate, and nobody in the HN thread could confirm either.
The interesting follow-up isn't whether Siri can use Claude. It's whether the Apple Intelligence team can ship a version in the next two years that makes users not want to switch.
Sources:
- MacRumors: Apple's Siri AI Can Be Swapped Out for Claude, ChatGPT, Code Shows (2026-09-14)
- Hacker News thread item 49695409 — 83 points on the front page, 25 comments