Siri AI

At WWDC 2026, on what became Tim Cook's farewell keynote, Apple finally shipped the conversational Siri it first promised in 2024 and delayed through 2025: a standalone app that chains multi-step actions across apps, reads your screen and your inbox, searches the live web, and carries one conversation across your devices. It is the biggest Siri overhaul in about fifteen years. The detail that matters is what's underneath it.

We use none of the models that Google deploys to their customers, nor do we use the infrastructure. — Craig Federighi, Apple

Apple's assistant now runs on three tiers. The lightest queries stay on the phone, on Apple's own models. The heaviest reasoning goes to a custom Gemini model — roughly 1.2 trillion parameters — that Apple licenses from Google for a reported $1 billion a year and runs on Google Cloud. Apple's on-device models, the third-generation Apple Foundation Models, were themselves trained on Gemini's outputs, a standard distillation move that copies a stronger model's behavior into a smaller one. The company that built its brand on 'no one but Apple touches your AI' has Google's brain in the stack.

Apple didn't hide it — that's the strange part. The Gemini collaboration was announced on stage as an upgrade, and in the same week Craig Federighi told the press Apple uses 'none of the models that Google deploys to their customers, nor the infrastructure.' Both things are true: it isn't the consumer Gemini you can sign up for, it's a private custom one, and Apple says the queries are stateless with Google contractually barred from training on Apple user data. But the lawyerly precision of that sentence is the tell. The real concession isn't about privacy mechanics; it's that the last big-tech holdout has admitted it could not field a competitive frontier model on the timeline, and rented one — the same move Amazon made with Anthropic and Samsung with Google.

Ship date is fall 2026, on iOS, iPadOS and macOS 27, to a base north of a billion active iPhones. The largest consumer-AI rollout of the cycle will, for its hardest questions, be running on Google's model.