Google's cloud chief Thomas Kurian dropped the Gemini-Siri confirmation mid-keynote at Cloud Next yesterday, almost casually — like it was old news. And in a way, it is. The deal's been public since January. But what most people missed is that the first phase already shipped weeks ago in iOS 26.4, and it's quietly changing how Siri interacts with every app on your phone.
What Google Confirmed at Cloud Next
The financial scaffolding: Apple pays Google roughly $1 billion per year for access to a custom 1.2-trillion-parameter Gemini model — mixture-of-experts architecture, tuned specifically for Siri and Apple Intelligence workloads. What Kurian confirmed yesterday was the timeline and scope. Phase 1 is live now. Phase 2 lands with iOS 27 in September. And the collaboration goes deeper than licensing weights — Google's team is actively optimizing inference for Apple's Private Cloud Compute silicon.
The subtext is hard to miss. Apple — a company that famously builds everything in-house, from its own chips to its own programming languages — looked at the state of its language models and decided the fastest path forward was writing Google a billion-dollar check. That tells you something uncomfortable about where Apple Intelligence stood relative to the frontier. Tim Cook reframed it as a privacy play, not a capability gap. Make of that what you will.
Siri Sees Your Screen Now — and That Rewrites the Rules
The Phase 1 capability that should have every iOS developer's attention is on-screen context awareness. Since iOS 26.4, Siri can read and reason about whatever is currently displayed on the user's device. Not just your app's declared intents or shortcuts — the actual pixels and text on screen.
Someone looking at a restaurant listing in Safari? Siri makes the reservation without them copying the name or address. Flight confirmation email open in Mail? Siri adds it to Calendar and sets a departure reminder. No app-switching, no manual data transfer, no friction. The Gemini model's multimodal reasoning makes this work in a way that Apple's own models couldn't pull off reliably.
For developers, this is a slow-motion earthquake. Your app's UI is now an input surface for the system assistant. Every visible text element, every label, every data point on screen becomes something Siri can potentially reason about and act on. If you've been cutting corners on accessibility labels, using images where you should have text, or shipping unlabeled buttons — those aren't just accessibility issues anymore. They're discoverability issues for the most powerful feature Apple has shipped in years.
The action-chaining is the other big unlock. Siri handles up to 10 sequential actions from a single natural language request. "Book me on the next flight to New York, add it to my calendar, and text Sarah my arrival time" runs as one workflow — not three separate commands with confirmation dialogs jammed between each step. This orchestration depends on App Intents integration, and if you haven't adopted the framework yet, the gap between your app and your competitors' is about to get very visible.
The Free Lunch (Sort Of)
Here's the part developers will like: if you've already implemented App Intents or SiriKit domains, your existing integrations get smarter automatically. The Gemini upgrade operates at the model layer, not the API layer. Your shortcuts work. Your intent handlers work. They just understand more nuance.
SiriKit now covers 340+ intent categories, up from 120 in iOS 25. One genuinely new developer-facing addition: intent handlers now receive a processingTier callback indicating whether the query was resolved on-device or in the cloud, so you can adjust response detail accordingly.
September's Phase 2 Is Where Apple Gets Ambitious
Phase 2 ships with iOS 27, and this is where the strategy takes a sharp turn. Full conversational Siri arrives — maintaining context across 20+ exchange dialogues, handling complex multi-app automation, and building personalized behavior profiles over time.
But the real headline is the model marketplace. iOS 27 will let users choose any AI model from the App Store as Siri's brain. That sentence deserves a second read. Apple is building an open inference layer where third-party model providers can compete to power the system assistant. You could ship a model fine-tuned for medical terminology, legal research, or creative writing — and users could slot it in as their default Siri backend.
Whether this actually works is a separate question. App Store review for AI models raises dozens of unresolved issues around safety, consistency, and the user experience fracturing across a thousand different "Siris." But the ambition is revealing: Apple wants to own the platform for AI, not be the AI. It's the App Store playbook applied to intelligence itself.
Apple's Privacy Play Makes This Possible
The architecture underneath deserves a mention because it's what makes the whole thing politically viable. Google licensed frozen model weights to Apple — not a live API. Apple runs all inference within Private Cloud Compute on custom Apple Silicon servers running a hardened OS built specifically for private AI workloads. Google gets zero logs, zero prompts, zero user telemetry. It's a clean split: Google supplies the intelligence, Apple supplies the trust boundary.
Privacy advocates have been cautiously positive, though independent verification of PCC remains limited. The "invisible catch," as TheStreet put it, is that you're trusting Apple's claim that the isolation works as described.
Apple just admitted it can't build the best language model. But it might have built the best container to run one in. For developers on the platform, that distinction could end up mattering more than the model itself.