Apple’s Siri Can Be Swapped for Claude or ChatGPT, Hidden Code Reveals

Apple built Siri to be replaceable. That’s the takeaway from a fresh round of code digging into iOS 27 and macOS Golden Gate, and it’s a stranger admission for a company that has spent a decade selling the tightness of its own ecosystem as a feature.

The discovery comes from a developer going by pdfu, a regular source of pre-release iOS leaks, who found two private frameworks buried in the current release candidates. Neither is visible to ordinary users. Both point toward a version of Siri where Apple’s own model is one option among several, not the only one.

Two Mechanisms, Two Levels of Depth

The shallower of the two is called Model Delegation, built on Apple’s existing App Intents system. In a demo video posted to X, pdfu pulls up the “Search or Ask” bar on a Mac, selects Claude from a contextual “Ask…” menu, and requests a reminder. Claude parses the request in natural language, then hands the actual system action back to Siri, which creates the entry in the Reminders app.

That handoff boundary is the whole design. Claude can reason over a request and produce something Siri can’t natively generate; the same demo shows it producing a CSV file, but the moment a task needs to touch an Apple system feature, control routes back to Siri’s own execution layer. Functionally, it slots into the same UI position ChatGPT already occupies as a built-in extension.

The second mechanism goes further. Called Inference Provider, part of a broader Model Manager Services framework, it lets an external model take over Siri’s server-side role outright. In pdfu’s second demo, GPT-5.6 Terra receives Apple’s own Siri planner prompt and tool definitions directly, then executes tasks, searching device email, pulling specific information out of the inbox, and sending iMessages using that data, through Siri’s interface and voice, as if it were Siri itself.

The distinction matters. One mechanism adds a third-party model as a helper Siri can call on. The other replaces Siri’s decision-making core with someone else’s model while keeping Apple’s interface as the shell.

Not Live, and Not Confirmed

None of this is available to users right now. The “Ask…” menu in the current macOS Golden Gate release candidate, set for release on September 14, lists only Apple’s built-in ChatGPT extension. Claude does not appear as a user-selectable option, and Apple has not opened the Model Delegation entitlement to any third-party developer, Anthropic included.

The Regulatory Backdrop

The likeliest explanation for why this exists at all isn’t a change of philosophy at Apple. It’s the European Union’s Digital Markets Act, which requires the company to give third-party developers access to hardware and software features it previously kept closed. The law specifically named Siri and Google’s Gemini as targets for interoperability requirements, aimed at preventing device makers from using default-app positioning to box out competitors.

Apple has already responded to that pressure by withholding its new Siri AI features from EU users entirely when iOS 27 rolls out, rather than opening access on the terms Brussels wants. Code built for future compliance, sitting dormant in a build that currently excludes the market driving that compliance, fits a company managing a regulatory deadline rather than a company embracing openness.

What It Means for OpenAI and Anthropic

If Apple does eventually open these frameworks to outside developers, the stakes are different for each company involved. ChatGPT already has a foothold as Siri’s built-in fallback option; this would extend an existing relationship rather than start one. For Anthropic, appearing in Claude’s leaked demo without a live toggle is a signal of technical readiness, not a commercial agreement; nothing in the code confirms a deal is in place.

Both companies have reason to want the access regardless. iPhone represents one of the largest concentrated smartphone user bases in a single market, and both OpenAI and Anthropic are pursuing IPOs in an environment where user reach and distribution deals factor directly into valuation conversations. Landing as a default or near-default option on hundreds of millions of devices would be worth more than the integration itself; it would be proof of relevance in a market both companies are still fighting to define.

For now, that’s a possibility written into code, not a product anyone can use.