Apple engineered its rebuilt Siri to hand control to outside AI models at a far deeper level than anything it has shipped publicly. That is the conclusion researcher “pdfu” reached after digging through private frameworks inside iOS 27 and the macOS 27 “Golden Gate” release, work first reported by MacRumors.
Two separate mechanisms surfaced in the code. The first, called Model Delegation, lets a model such as Claude sit inside Siri the way the shipped ChatGPT extension already does: a user picks it from an “Ask…” menu, poses a request in plain language, and the assistant either answers directly or passes system tasks, like creating a reminder, back to Siri to execute. A demo pdfu posted showed exactly that handoff, plus a case where the outside model built a CSV file, something Siri cannot do on its own.
The second mechanism goes further. Through a component called Model Manager Services, Apple can apparently swap out its own server-side Siri model entirely and route Siri’s planning prompts and tool definitions to a replacement, in the demonstrated case GPT-5.6. The substitute model gets the same system access Siri’s native model would have, including personal data needed to complete a task, then hands its answer back for Siri to speak in its own voice and interface. In one clip, the swapped-in model searched a user’s email, summarized the findings, and sent a Messages text to a contact, with the exchange visible afterward in OpenAI’s own developer console.
Neither path is available to Apple users today. The Ask menu currently only supports the built-in ChatGPT extension, and that is limited to the macOS 27 Golden Gate release candidate, which MacRumors describes as effectively the shipping build. Apple has not opened the underlying Model Delegation entitlement to outside developers, and none of this is surfaced in any released setting.
The timing points to regulation as much as product strategy. The European Union’s Digital Markets Act requires Apple to give rival services meaningful access to iOS features it reserves for its own software, and the European Commission has said that obligation covers Siri specifically. Building a formal delegation and inference-substitution layer, rather than a one-off ChatGPT integration, reads as Apple preparing compliance infrastructure ahead of enforcement rather than reacting to a single complaint.
It also reframes what “choosing an AI model” could mean on Apple hardware. Most model-switching features elsewhere, in browsers or on Android, change which chatbot answers a prompt in a chat window. Apple’s architecture, if activated, would let an outside model operate as the reasoning layer behind system-level actions across Reminders, Mail, and Messages, while Apple keeps the interface, the voice, and the data-handling boundary. That is a materially different competitive position for Anthropic and OpenAI than winning a chatbot preference toggle.
Because the code is unreleased and undocumented, Apple has not said which models would qualify, what data-sharing terms would apply, or when either mechanism might reach users. Developers building Apple-adjacent AI features should treat this as a signal to plan for a future deep-integration entitlement, not as a shipped API, and watch the next Golden Gate release candidates for signs the ChatGPT-only limit is lifted.
Reporting by Tim Hardwick for MacRumors, published September 14, 2026.
