Apple has spent fifteen years telling you that Siri is getting smarter. It wasn't. And now, in the most quietly significant admission in Silicon Valley history, the company is starting over — using a competitor's engine to do it.
Sources confirmed this week that iOS 26.4, expected in late March 2026, will ship with a fundamentally rebuilt Siri powered by Google's Gemini model at 1.2 trillion parameters. The original Siri architecture — built on a 2011 acquisition of a Stanford spinout called SRI — will be retired entirely for on-device intelligence tasks.
What Went Wrong With the Old Siri
Siri was genuinely impressive when it launched in 2011. It was years ahead of Google Assistant, and it made the iPhone feel like science fiction. But Apple built it on a narrow intent-classification stack that was fast and private but fundamentally incapable of reasoning.
Every other major AI company pivoted toward large language models between 2020 and 2023. Apple's AI team fought internally for resources, lost repeatedly to hardware priorities, and produced Apple Intelligence in 2024 — a product so disappointing that Phil Schiller reportedly refused to demo it at the internal reveal. The gap between Siri and ChatGPT became embarrassing in real-world use.
The tipping point came late in 2025, when an internal Apple study found that 67% of users who owned an iPhone 16 had never used Siri for anything more complex than setting a timer. The assistant they had spent billions building was being used as a glorified clock.
What Gemini 1.2T Actually Brings
The Gemini model Apple licensed isn't the consumer version you use in Google's apps. It's a fine-tuned derivative of Gemini Ultra trained specifically on Apple device context — app states, system calls, calendar permissions, health data patterns — with strict on-device processing constraints Apple negotiated into the agreement.
At 1.2 trillion parameters, it's roughly six times the scale of GPT-4 as publicly described, though parameter counts are an increasingly unreliable proxy for capability. What matters is what it can do: multi-step reasoning, ambiguous query resolution, and crucially, cross-app context that old Siri never achieved.
Ask the new Siri to "find the restaurant where I had dinner with Marcus last April and book the same table for Friday," and it will cross-reference your Messages, Maps history, calendar, and OpenTable integration in a single pass. Old Siri would have asked you to clarify what you meant by "Marcus."
On-Screen Awareness: The Feature That Changes Everything
The headline feature of the rebuilt Siri is what Apple is calling "on-screen awareness." The model can read, interpret, and act on whatever is currently displayed — not just within Apple apps, but across any app that has implemented the new SiriKit overlay API.
In a demo shared with developers this week, Siri was shown reading a long email thread in Gmail, summarizing it, drafting a reply in the user's writing style, and scheduling a follow-up reminder — all from a single voice command, without the user leaving the Gmail app.
This is the feature that Microsoft Copilot promised in 2023 and largely failed to deliver on mobile. Apple, with its hardware-software integration advantages and on-device compute through the M4 Neural Engine, may actually be positioned to make it work reliably.
The Google Deal — And What Apple Gave Up
The partnership is built on a revenue-sharing agreement that sources describe as "highly unusual" — Apple pays per inference for cloud-routed queries while keeping on-device processing free. Google gets a guaranteed distribution channel for Gemini to over 1.4 billion active Apple devices, a number no other distribution deal comes close to matching.
Privacy advocates have raised immediate concerns. Apple has historically been the most privacy-forward of the major tech companies, and handing inference work to Google — even with contractual guardrails — is a philosophical reversal. Apple's official statement says queries routed to cloud are anonymized and not used for training. Privacy researchers are, predictably, skeptical.
The deal expires in 2028, with Apple reportedly building its own LLM internally as a longer-term replacement. But for now, if you want a Siri that works like it was supposed to back in 2011, it runs on Google's servers.
iOS 26.4 Launch and What to Expect
The rollout will be staged. On launch day, on-screen awareness will be limited to first-party Apple apps plus 14 third-party partners including Gmail, Slack, and Spotify. The full SiriKit overlay API opens to all developers in Q2 2026.
On-device processing handles queries that don't require real-time data or complex reasoning — estimated at roughly 70% of everyday Siri interactions. Cloud routing kicks in for everything else, with a subtle visual indicator users can see. Apple says the typical response latency for cloud-routed queries is 280 milliseconds — faster than any third-party AI assistant currently available on iOS.
What This Means For You
If you own an iPhone and gave up on Siri years ago, iOS 26.4 is worth revisiting. The feature that will change your daily use most immediately is cross-app action — the ability to give Siri a multi-step task across apps you already use and have it complete it without you watching.
If you're an Android user, this is competitive pressure on Google to accelerate what Gemini can do natively on Pixel and Samsung devices. A rising Siri is good for everyone who uses a phone.
And if you're watching the AI industry: Apple just signaled that building foundation models is harder than it looks, even for one of the most valuable companies in human history. The next two years of AI will be about distribution, not just capability — and Apple just secured the most valuable distribution in the world for Google's model.