Apple's Siri AI Overhaul: The Gemini Deal in 2026

TL;DR

• Apple is paying Google roughly $1 billion per year for access to a custom 1.2-trillion-parameter Gemini model that will power a completely redesigned Siri, expected to ship with iOS 27 later in 2026.¹²
• As of March 25, 2026, Apple can now distill the full Gemini model into smaller, task-specific models built to run entirely on-device — no internet required.³⁴
• All cloud-based AI processing runs on Apple's Private Cloud Compute (PCC) infrastructure, using Apple Silicon servers with stateless, ephemeral computation — meaning user data is never stored or accessible, even to Apple.⁵
• The new Siri, codenamed "Project Campos," will feature on-screen awareness, cross-app reasoning, conversational memory, and a chatbot-style interface rivaling ChatGPT and Claude.⁶⁷
• WWDC 2026 on June 8 is expected to be the flagship reveal, with the full rollout alongside iOS 27.⁷

What You'll Learn

1. What Apple's Gemini deal actually includes — the technical scope, financial terms, and privacy architecture.
2. How model distillation works and why it matters for on-device AI.
3. What the redesigned Siri will be capable of, based on confirmed reports and leaks.
4. When these changes ship and what developers should watch for.
5. Why this partnership reshapes the competitive dynamics of the AI assistant market.

The Partnership: What Apple and Google Actually Agreed To

Apple Intelligence is the brand name for Apple's suite of AI features across iPhone, iPad, and Mac. Since its initial launch, Apple Intelligence has relied on a compact on-device foundation model — approximately 3 billion parameters — for tasks like text summarization and notification prioritization, alongside a larger server-side model for heavier workloads.²⁸

That changed on January 12, 2026, when Apple and Google issued a joint statement announcing a multi-year collaboration. Under the deal, the next generation of Apple Foundation Models will be built on Google's Gemini technology and cloud infrastructure. These models will power future Apple Intelligence features, including a dramatically more capable Siri.¹⁹

Financial Terms

Apple is reportedly committing approximately $1 billion per year for access to a custom Gemini model with an estimated 1.2 trillion parameters — a roughly 400x increase in model scale compared to Apple's previous ~3-billion-parameter on-device architecture.²¹⁰⁸ Neither company has officially confirmed the exact price, but multiple independent reports cite this figure.

What Gemini Handles

The custom Gemini system is responsible for Siri's summarizer and planner functions — the components that help the assistant synthesize information from multiple sources and decide how to execute multi-step tasks.¹ Crucially, Gemini's role is "white-labeled": there is no Google branding visible to end users. From a user's perspective, this is still Siri.²

How Distillation Brings Gemini to Your iPhone

On March 25, 2026, new reports confirmed that Apple now has the ability to distill Google's full Gemini model into smaller, specialized models designed to run on Apple devices without an internet connection.³⁴

What Is Model Distillation?

Model distillation (also called knowledge distillation) is a technique where a large "teacher" model transfers its learned capabilities to a smaller "student" model. Instead of training the student from scratch on raw data, you feed it the teacher's probability distributions — its reasoning patterns, confidence levels, and decision-making processes.⁸

The result is a compact model that retains much of the teacher's performance at a fraction of the computational cost. This is how companies like Apple can take a 1.2-trillion-parameter cloud model and produce efficient variants that run on a phone's Neural Engine.

How Apple's Distillation Pipeline Works

According to reports, Google gave Apple "complete access" to the Gemini model running in dedicated data centers.³ Apple's process works as follows:

1. Task generation: Apple asks the full Gemini model to perform a series of representative tasks — answering questions, summarizing documents, planning multi-step actions.
2. Reasoning capture: For each task, Gemini produces high-quality results along with its internal reasoning chain.
3. Student training: Apple feeds both the answers and reasoning data to smaller models, which learn to replicate Gemini's internal computations.³
4. On-device optimization: The resulting distilled models are tuned for Apple Silicon's Neural Engine, enabling local inference without cloud connectivity.

This approach lets Apple maintain a privacy-first architecture while benefiting from Google's massive training investment. The distilled models can handle tasks like text summarization, contextual understanding, and intent classification entirely on-device.

For a deeper look at running AI models locally, see our guide on running LLMs locally and how to save costs with small LLMs.

Private Cloud Compute: The Privacy Architecture

When on-device models are not sufficient — for complex queries or tasks requiring the full Gemini model — Siri falls back to Apple's Private Cloud Compute (PCC) infrastructure.⁵

PCC is not a typical cloud service. It is built on custom Apple Silicon servers running a hardened operating system specifically designed for private AI processing. The key security guarantees include:

This hybrid architecture — distilled on-device models for common tasks, PCC for complex ones — means that even when your query does touch the cloud, your data receives stronger privacy protections than any competing AI assistant offers today.

For more on building privacy-respecting AI systems, see our post on building private AI models with open-source LLMs.

What the New Siri Will Actually Do

The redesigned Siri, developed under the internal codename "Project Campos," represents Apple's most ambitious assistant overhaul since Siri's original 2011 launch.⁶

Confirmed Capabilities

Based on confirmed reports and Apple's own statements, the new Siri will include:

• On-screen awareness: Siri will be able to see and understand what is currently displayed on your screen — a flight confirmation in an email, a photo in Messages, a product on a webpage — and take contextual actions based on that content.⁷
• Cross-app reasoning: Rather than being limited to single-app commands, Siri will chain actions across multiple apps. For example, extracting a date from an email, creating a calendar event, and sending a confirmation message — all from a single request.⁷
• Conversational memory: Siri will remember previous conversations and use that context to provide more personalized, coherent responses over time.²
• Proactive suggestions: The assistant will anticipate needs — suggesting you leave early based on traffic data before an airport pickup, or surfacing relevant documents before a meeting.²
• Ask Siri toggle: A new UI element appearing across the app ecosystem where users can highlight content and ask Siri for contextual actions — summarizing, translating, searching, or acting on the selected content.⁷

The Chatbot Mode

Apple is also developing a standalone chatbot version of Siri that will compete directly with ChatGPT, Gemini, and Claude. This mode will handle extended conversations, document analysis, creative writing, emotional support queries, and complex multi-step tasks like booking travel.⁶⁷

A preview of the chatbot Siri could be shown at WWDC 2026 on June 8.⁷

Timeline and What to Watch For

Apple has delayed the "more personalized Siri" multiple times over the past year, leading to declining consumer perception of Apple Intelligence.¹¹ The Gemini partnership is widely seen as Apple's strategy to close the gap rapidly rather than building frontier-scale models entirely in-house.

The Competitive Landscape: Why This Matters

Scale of Impact

Apple reported 2.5 billion active devices worldwide as of January 2026, up from 2.35 billion a year earlier.¹² Siri is active on the vast majority of these devices — Apple last disclosed over 500 million Siri-enabled devices in 2018, and the number has grown substantially since.¹¹ When the Gemini-powered Siri ships, it will instantly become the most widely distributed AI assistant with frontier-model capabilities — reaching users who have never interacted with ChatGPT or Claude.

The Privacy Differentiator

Unlike competing assistants that process queries on general-purpose cloud infrastructure, Apple's hybrid approach (on-device distilled models + PCC) offers a unique privacy proposition. No other major AI assistant combines frontier-model performance with stateless cloud computation and hardware-verified privacy guarantees.

Developer Implications

For developers building on Apple platforms, this shift means:

• SiriKit and App Intents will likely expand significantly to support cross-app reasoning and on-screen awareness.
• On-device model APIs may become available for developers to leverage Apple's distilled models in their own apps.
• Privacy-first AI patterns that Apple enforces will set expectations for how users want AI features to handle their data.

For context on how AI is reshaping development workflows, see our guide on the future of LLMs and fine-tuning.

Key Takeaways

The Apple-Google Gemini deal is not just another partnership announcement. It is the mechanism through which the world's most privacy-focused consumer tech company gains access to frontier AI capabilities — and through distillation, brings those capabilities directly to the device in your pocket.

The three things to watch: first, how much on-device capability the distilled models actually deliver when iOS 27 ships; second, whether PCC's privacy guarantees hold up under independent security research; and third, how Apple opens these capabilities to third-party developers at WWDC.

If you're interested in the broader AI hardware race powering these models, read our deep dive on the custom AI chip race in 2026.