Apple is paying Google roughly $1 billion per year to make Siri competent. That is the simplest summary of the multi-year partnership announced in January 2026, under which the next generation of Apple Foundation Models will run on Google’s Gemini architecture and cloud infrastructure. The first results ship with iOS 26.4, rolling out to compatible iPhones by late March 2026: on-screen context awareness, cross-app task chains, and conversational memory that persists across turns. Siri goes from a voice command parser to something that starts to resemble an actual AI agent.
The strategic logic is blunt. Apple spent two years building its own foundation models after the ChatGPT launch and fell visibly behind. Rather than ship another incremental update, Apple wrote a check to Google and got access to models that already work. The deal is white-labeled: no Google branding appears anywhere in the user experience. From the outside, this is still Siri. From the inside, it is Gemini wearing Siri’s skin.
What the $1 Billion Actually Buys
The Apple-Google agreement covers three layers. First, Apple gets access to Gemini’s base models to train its own Apple Foundation Models. This is not a simple API call to Google’s servers; Apple is using Gemini’s architecture as the foundation for models it fine-tunes and deploys on its own infrastructure. Second, Apple gets access to Google’s cloud compute for the heavier inference tasks that cannot run on-device. Third, the deal includes ongoing model updates, meaning Apple’s capabilities will improve as Google ships newer Gemini versions.
Neither company has confirmed the exact price. Multiple reports cite roughly $1 billion annually, which would make this one of the largest AI licensing deals in history. For context, that is roughly what Apple was paying Google for the Safari default search deal before it ballooned to $20 billion per year. The AI deal is a fraction of that, but it signals how seriously Apple takes the gap between Siri and the competition.
The financial structure also reveals Apple’s bet: it is cheaper and faster to license best-in-class models than to build them. Apple’s own AI research team has published strong work on efficient on-device models, but the Gemini deal acknowledges that for complex reasoning, multi-step planning, and vision tasks, Google is ahead.
On-Screen Awareness: Siri Can Finally See
The headline feature in iOS 26.4 is on-screen context awareness. Siri can now interpret the content displayed on your screen and act on it without you having to copy, paste, or explain what you are looking at.
The practical examples matter more than the marketing language. If you have a restaurant open in Safari, you can say “make a reservation for two tonight” and Siri will read the restaurant name, find its booking system, and attempt the reservation. If you are looking at a flight confirmation email, saying “add this to my calendar and set a reminder two hours before departure” triggers Siri to extract the flight details, create the calendar event, and configure the reminder, all from a single utterance.
This is not optical character recognition bolted onto a voice assistant. The Gemini-powered vision system understands semantic context: it knows that the phone number on a restaurant page is different from a phone number in a spam email. It can distinguish between a photo you want to share and a screenshot of a conversation you are reading. Apple calls this “screen understanding,” and it represents the jump from keyword matching to genuine comprehension.
The limitation is worth noting. On-screen awareness in iOS 26.4 works only within Apple’s own apps and a set of whitelisted third-party apps. Full third-party support, where Siri can read and act on the screen of any app, is expected in iOS 27 at the earliest.
Cross-App Actions and Task Chains
The second major capability is multi-step task execution. Siri can now chain up to 10 sequential actions from a single natural language request, working across multiple apps without returning control to the user between steps.
A concrete example: “Find the cheapest flight to Berlin next Friday, book it with my saved payment method, add it to my calendar, and text Sarah the itinerary.” That is four distinct actions spanning Safari, Wallet, Calendar, and Messages. In previous iOS versions, each step required a separate Siri invocation. Now, Siri plans the sequence, executes each step, and reports back when finished.
The task chain architecture uses a planner-executor pattern that will look familiar to anyone who has built with LangChain or CrewAI. Siri’s planner decomposes the request into sub-tasks, determines the execution order, identifies which apps need to be invoked, and passes context between steps. If a step fails (the flight is sold out, the payment method is expired), Siri reports the failure and asks for guidance rather than silently stopping.
Conversational memory is the glue that makes this work. Siri now maintains context across up to 50 turns in a single session. You can ask follow-up questions, refine instructions, or change parameters without restating the full request. “Actually, make that Saturday instead of Friday” works because Siri remembers the entire chain it just planned.
The Privacy Architecture: Why It Matters More Than the Model
Apple’s pitch for this deal rests on a specific technical claim: your data never touches Google’s servers directly. The privacy architecture uses a three-tier system.
Tier 1: On-device processing. Simple tasks (setting timers, playing music, basic queries) run entirely on the iPhone’s Neural Engine using Apple’s smaller on-device models. No data leaves the phone.
Tier 2: Private Cloud Compute. More complex tasks that exceed on-device capability are processed on Apple’s own cloud infrastructure using custom Apple Silicon servers. Apple has published the Private Cloud Compute architecture in detail: data is encrypted in transit and at rest, processed in ephemeral compute nodes that retain no data after the request completes, and the code running on these servers is publicly auditable.
Tier 3: Google Cloud (anonymized). Only the most demanding queries, primarily those requiring Gemini’s full reasoning capability, are routed to Google’s infrastructure. Before any data reaches Google, Apple strips personally identifiable information through a privacy buffer layer. Google cannot associate queries with specific Apple users or devices.
This matters because it addresses the core objection enterprises and privacy-conscious users have with cloud AI: who sees my data? Apple’s answer is layered defense. Most queries never leave your phone. The ones that do are processed on Apple hardware first. Only the hardest problems touch Google, and even then, only in anonymized form.
What Got Delayed: The iOS 26.5 and iOS 27 Roadmap
Not everything made the cut. Apple originally planned to include the full Gemini-powered Siri overhaul in iOS 26.4, but internal challenges forced a phased rollout.
iOS 26.4 (March 2026): On-screen awareness in Apple apps, basic cross-app task chains, conversational memory, improved natural language understanding.
iOS 26.5 (May 2026): Extended third-party app support for on-screen awareness, deeper Shortcuts integration, proactive suggestions based on usage patterns.
iOS 27 (September 2026): Full agentic mode where Siri can operate autonomously on multi-step tasks with minimal user intervention, third-party agent plugins, and persistent memory across sessions.
The delay is strategically interesting. Apple’s own UX research found that users prefer transparent, predictable AI agents over powerful but opaque ones. Shipping a phased rollout gives users time to build trust with Siri’s new capabilities before Apple unlocks full autonomous operation. Whether that is genuine user-centered design or post-hoc rationalization for missing deadlines depends on your level of cynicism.
What This Means for the AI Agent Market
The Apple-Google deal reshapes the competitive landscape in three ways.
Consumer AI agents are real now. When Siri becomes a genuine AI agent on 1.5 billion active Apple devices, the concept of “AI agents” stops being an enterprise buzzword and becomes something your parents use. This is the mass-market inflection point the industry has been waiting for. Every competitor, from Samsung’s Bixby to Amazon’s Alexa, will face pressure to match these capabilities.
The model layer is commoditizing. Apple, one of the wealthiest companies on earth, decided it was better to license Google’s models than build its own. That sends a clear signal: the value is not in the foundation model itself but in the product experience, privacy architecture, and distribution built on top of it. Startups building “better models” should take note.
Privacy becomes a feature, not a constraint. Apple’s three-tier architecture proves you can build a powerful AI agent that respects user privacy. The “you have to choose between capability and privacy” argument just lost its strongest example. Expect European regulators to point to Apple’s approach when drafting AI agent privacy standards.
For enterprise buyers evaluating AI agent platforms, the Apple deal validates a key architectural pattern: use the best available models (even from a competitor), wrap them in your own privacy and security layer, and own the user experience. That is a template any company can follow.
Frequently Asked Questions
How much is Apple paying Google for the Gemini Siri deal?
Multiple reports estimate the deal at roughly $1 billion per year, though neither Apple nor Google has confirmed the exact figure. The multi-year agreement gives Apple access to Gemini’s base models, cloud compute infrastructure, and ongoing model updates for powering the next generation of Siri.
Does Apple send my Siri data to Google?
Not directly. Apple uses a three-tier privacy architecture. Simple queries run entirely on your iPhone. More complex tasks are processed on Apple’s Private Cloud Compute servers. Only the most demanding queries reach Google’s infrastructure, and Apple strips personally identifiable information through a privacy buffer layer before any data leaves Apple’s systems.
What new Siri features come with iOS 26.4?
iOS 26.4 introduces on-screen context awareness (Siri can see and understand what is displayed on your screen), cross-app task chains (up to 10 sequential actions from a single request), and conversational memory (Siri maintains context across up to 50 turns). Some features like full third-party app support are delayed to iOS 26.5 and iOS 27.
Is Siri now as good as ChatGPT or Google Assistant?
Siri’s reasoning capabilities are now powered by Gemini, which puts it in the same model class as Google Assistant. However, Siri’s initial iOS 26.4 release has limited third-party app support compared to Google Assistant’s broader integration. The full agentic capabilities, including autonomous multi-step operation and third-party plugins, are expected in iOS 27 in September 2026.
Will the Gemini-powered Siri work on older iPhones?
The Gemini-powered Siri features require Apple Intelligence, which is limited to iPhone 15 Pro and later models. On-device processing tasks require the A17 Pro chip or newer. Older iPhones running iOS 26.4 will receive other updates but not the new AI agent capabilities.