Table of Contents
- What the Deal Actually Covers
- What the New Siri Can Do That the Old One Couldn’t
- The Tiered Architecture Trick
- Google Now Owns Two Queries on Every iPhone
- The Privacy Story Needs an Asterisk
- When It Ships and What to Do About It
- FAQ
Apple has spent fifteen years building a brand on one principle: don’t outsource the parts that matter. CPUs, modems, maps, even email — buy them in early, then replace them with in-house versions the moment the technology is ready. The new Siri is the exception. Apple just signed a multi-year deal to rent its AI brain from Google for roughly a billion dollars a year. The Siri shipping in iOS 26.4 today and the “full conversational” Siri arriving with iOS 27 in September are both powered by a custom 1.2-trillion-parameter Gemini model.
That’s not a partnership. That’s an admission. Apple’s AI strategy is no longer “build it ourselves.” It’s “buy the best one available and put it inside Apple’s distribution and privacy layer.” The implications are bigger than the Siri capability list, and most of the coverage missed them.
What the Deal Actually Covers
Apple and Google announced the partnership jointly on January 12, 2026. The deal: Apple pays Google approximately $1 billion per year for a custom 1.2-trillion-parameter Gemini model, optimized for mobile interactions and adapted to Apple’s privacy architecture. The next generation of Apple Foundation Models — the family of models that power Apple Intelligence across iPhone, iPad, and Mac — will be built on Gemini and Google Cloud infrastructure rather than the in-house models Apple shipped in 2025.
This is the same Apple that held off on a usable AI Siri for two years rather than ship a version that didn’t meet its quality bar. The conclusion of that two-year delay is now public: Apple decided it was faster to license the gap than close it.
What the New Siri Can Do That the Old One Couldn’t
The capability jump is real. Phase 1, already shipping in iOS 26.4, brings on-screen awareness — Siri can read what’s on your screen and act on it without you describing it. Pull a flight number out of an email. Extract a reservation time from a Messages thread. Reference a chart on screen during a question. Phase 2, arriving in September with iOS 27 and the iPhone 18, adds full conversational dialogue across 20+ exchanges, the ability to chain multiple commands into a single request, and a standalone Siri app with persistent chat history that looks structurally like ChatGPT or Gemini.
The benchmarks Apple has shared are ambitious: 92% completion on complex multi-step tasks (up from 58% on the old Siri) and sub-0.5-second response latency on most queries. Early iOS 26.4 telemetry reports 67% daily engagement with the new Siri features, against 23% on the previous version. That gap is the kind of number Apple historically refuses to publish unless the answer is unambiguous.
The Tiered Architecture Trick
Apple is not just shoving every Siri query at Google’s servers. The architecture is tiered:
On-device. Simple tasks — set a timer, open an app, basic dictation, single-step commands — stay on the iPhone using Apple’s own small models. No data leaves.
Private Cloud Compute. Moderately complex tasks route to Apple’s encrypted server fleet, the same infrastructure Apple Intelligence has been using since 2025. Apple controls the hardware. Logs are scrubbed.
Custom Gemini. The heavy reasoning queries — the world-knowledge layer, the multi-step planning, the long conversational threads — hit the 1.2T-parameter Gemini variant. Reportedly hosted on infrastructure Apple controls, with the model weights licensed from Google rather than served from Google’s own data centers.
That last point is the part that lets Apple keep its privacy talking points alive. Whether it survives a serious audit is the part that didn’t get coverage.
Google Now Owns Two Queries on Every iPhone
Here is the angle most coverage missed. Google has been paying Apple roughly $20 billion a year for default search placement on Safari — the largest distribution payment in the history of software. Apple paying Google $1 billion a year for Gemini does not reverse that flow. Google still nets about $19 billion. What changes is what Google gets for it.
Before this deal, Google owned one query per iPhone: the search bar. Whatever you typed into Safari, Google saw and monetized. After this deal, Google owns the second query too: the AI conversation. Whatever you ask Siri once Phase 2 ships, a custom Gemini model processes. The model weights are Google’s. The training feedback loop, depending on contract terms, is at minimum partially Google’s.
For a 5% rebate of what Google was already paying, Google doubled the queries it sees from Apple’s user base. A billion-plus iPhones now route both their search intents and their conversational intents through Google models. That is not a story about an AI partnership. That is a story about the deepest distribution moat in tech getting deeper. Google’s $40 billion bet on Anthropic looks different in this light: Google is hedging on every model layer that matters, while collecting query data on the device layer it doesn’t own.
The Privacy Story Needs an Asterisk
Apple’s official line is that Apple Intelligence “continues to run on Apple devices and Private Cloud Compute, while maintaining Apple’s industry-leading privacy standards.” That is true for the on-device and PCC tiers. It is doing significant work for the Gemini tier.
The honest read: complex Siri queries — the kind people actually use AI assistants for — leave Apple’s pure on-device path. Whether the custom Gemini is hosted on Apple infrastructure (Apple’s claim), Google infrastructure (the simpler architecture), or some hybrid (more likely) determines whether your conversational Siri queries are seen by Google in any form. Apple has not published the contract terms. Independent privacy audits are not yet public.
For most users, the trade is fine. Better Siri is worth the privacy delta. Buyers who chose Apple specifically because of the on-device privacy story should at least understand what just changed.
When It Ships and What to Do About It
Phase 1 is live in iOS 26.4. Update and try it: hold down the side button, ask Siri to summarize an article on your screen, then ask it to draft a reply email referencing it. The cross-app awareness either works for you or it doesn’t. Phase 2 ships with iOS 27 in September, alongside the iPhone 18. The standalone Siri app, the 20-exchange conversational dialogue, and the chained-command capability all arrive then.
If you build product on top of Apple Intelligence APIs, the practical implication is that capability ramps significantly between now and September. Tasks that were impossible to ship on the 2025 Siri become viable on the Phase 2 model. Comparable AI assistant platforms like Claude Cowork and Perplexity Computer remain ahead on power-user workflows, but the distribution gap is now Apple’s to close.
If you are an enterprise privacy buyer specifically standardized on Apple because of the on-device guarantee, talk to legal before September. The 26.4 architecture is documented. The 27 architecture, with the standalone Siri app and conversational threads, is going to require new contracts and new disclosures. Better to read them before they ship than after.
FAQ
Is Apple paying Google or is Google paying Apple?
Both. Google pays Apple roughly $20 billion per year for default search placement on Safari and iOS. Apple now pays Google approximately $1 billion per year for the custom Gemini model powering Siri. Net flow remains heavily toward Apple, but Google’s strategic position has improved meaningfully.
What happens to Apple’s in-house AI models?
They continue to power on-device tasks and Private Cloud Compute workloads. Apple has not announced that it is abandoning its in-house frontier model effort. The Gemini deal covers the world-knowledge and complex-reasoning tier, not every layer.
Can I opt out of the Gemini integration?
You can disable Apple Intelligence entirely in iOS Settings. There is no current granular setting that keeps Apple Intelligence on while routing all heavy queries away from Gemini.
Does this mean ChatGPT integration is going away?
No. The existing ChatGPT integration in iOS, where Siri can hand off specific queries to ChatGPT with user permission, remains separate from the Gemini foundation. Users will be able to choose ChatGPT for individual handoff queries even on the Gemini-powered Siri.
Will Siri work the same on older iPhones?
Phase 1 features in iOS 26.4 work on iPhone 15 Pro and newer, plus M1 iPads and Macs. The full Phase 2 conversational Siri launching in September requires the neural-engine generation in iPhone 16 and later.
How does this compare to OpenAI’s distribution strategy?
OpenAI is pursuing the opposite playbook: own the consumer surface directly through ChatGPT and the recent Realtime API while licensing model access to enterprises. Apple’s deal gives Google what OpenAI is trying to build organically — frictionless access to a billion daily-active consumers. The GPT-5.5 launch and the OpenAI–Microsoft multicloud deal read differently when viewed against Google’s deepening iPhone footprint.