Apple Intelligence 2026: Inside Apple's Gemini-Powered Siri Overhaul
Published: July 26, 2026
Peak of Trending Editorial
Every Siri request more complicated than setting a timer now travels, at some point, through a language model Apple did not build. Apple Intelligence 2026 runs on three tiers of processing, and the top tier — the one handling research questions, multi-step planning, and anything close to real reasoning — is a custom 1.2-trillion-parameter Gemini model licensed from Google for a reported $1 billion a year. Apple's own cloud models, the ones the company spent three years building toward a fully in-house assistant, top out around 150 billion parameters. Nobody at Apple describes it that way on stage. The company's public language is "collaboration" and "the most capable foundation," phrasing built to make a licensing deal sound like a partnership announcement instead of an admission.
Most coverage of Apple Intelligence in 2026 treats the Gemini deal and the Siri relaunch as separate stories — a partnership headline in January, a feature-list roundup at WWDC in June. What gets lost is the sequence. Apple signed away its claim to owning the model layer, watched the reaction settle over the following months, and then replaced its chief executive for the first time in fifteen years three months after that. Covering those as unrelated news cycles misses the through-line running underneath both: an AI strategy that failed publicly enough to end a tenure that had already survived a flat iPhone cycle, a cool reception for the Vision Pro, and a canceled car project that burned a decade and roughly $10 billion.
What follows breaks down what Apple actually ships as Apple Intelligence in 2026, the mechanics of the Gemini arrangement underneath it, why the promised "more personal Siri" slipped past three iOS versions, and what the incoming Ternus era signals about whether a hardware-first bet on AI can work once the intelligence underneath belongs to somebody else.
- What Apple Intelligence Actually Runs On in 2026
- The Gemini Deal Apple Didn't Want to Announce This Way
- Why Siri's Rebuild Took Two Years Longer Than Promised
- The CEO Change That Wasn't Really About Retirement
- What Happens to ChatGPT Inside the New Siri
- Where Apple Intelligence Still Has an Actual Edge
- Who This Is For
- Verdict
What Apple Intelligence Actually Runs On in 2026
Apple Intelligence in 2026 splits every request across three separate layers before a user ever sees an answer. Simple commands — a timer, a text reply, a calendar lookup — stay entirely on the device, processed by Apple's own silicon and never touching a network connection. Moderately complex tasks route to Private Cloud Compute, Apple's own server infrastructure, where Apple's proprietary models handle the work under the same security guarantees as on-device processing. Anything closer to genuine reasoning — multi-step planning, open-ended research, a question Siri has to actually think through — now goes to a custom Gemini model that Google built specifically for Apple, confirmed jointly by both companies in a January 2026 statement covered by AppleInsider.
| Tier | What It Handles | Where It Runs |
|---|---|---|
| On-device | Timers, replies, quick lookups | Apple silicon, offline |
| Private Cloud Compute | Moderate tasks, personal context | Apple-owned servers |
| Gemini cloud | Reasoning, planning, research | Google infrastructure, Apple-controlled routing |
Google's Cloud chief described the arrangement from his side of the table, not Apple's, which is its own signal about who holds which piece of leverage in this deal.
"We're collaborating with Apple as their preferred cloud provider to develop the next generation of Apple Foundation Models based on Gemini technology," Thomas Kurian said, according to Google's own confirmation reported by AppleInsider.
150 billion parameters. That is where Apple's own cloud models stop, after roughly three years of internal development the company rarely quantified in public. The gap between that figure and Gemini's 1.2 trillion is not a rounding difference — it is the reason the routing table above exists at all, and the reason the hardest questions never touch Apple's own model.
The Gemini Deal Apple Didn't Want to Announce This Way
Reporting from Bloomberg's Mark Gurman put the arrangement's cost at roughly $1 billion annually, a figure repeated across outlets covering the January 12, 2026 joint announcement without Apple ever confirming the number directly. The custom model itself uses a mixture-of-experts architecture sized at roughly eight times Apple's own cloud models, built to handle summarization, planning, and natural-language understanding at a level Apple's internal teams had not reached on their own timeline.
Apple's own framing, released through its newsroom, called the choice the result of "careful evaluation" that found Google's technology offered "the most capable foundation" available. That is a defensible technical conclusion and also a costly one: Apple reportedly evaluated OpenAI's models during the same process before settling on Gemini, according to multiple outlets tracking the negotiation, which means the company priced out its options and picked the one that happened to belong to the search company it has spent a decade paying to remain the default.
None of this touches raw user data, and that distinction is the part of the deal Apple has been most careful to state plainly. Gemini's reasoning happens through Private Cloud Compute rather than on Google's own servers, and Apple has said queries routed through the arrangement are not retained or used to train Google's models. Users never see a Gemini interface, a Gemini logo, or a Gemini prompt box. They see Siri, asking a question, getting an answer, with the machinery underneath deliberately invisible.
Why Siri's Rebuild Took Two Years Longer Than Promised
Apple promised a more personal, context-aware Siri alongside iOS 18 in 2024. That version never shipped on schedule, slipped toward iOS 26.4 by early 2026, and then folded entirely into the iOS 27 release arriving with the iPhone lineup this September — a three-version delay that Computerworld's reporting on Apple's AI strategy traces directly to Apple's internal models falling short of what the company had already announced publicly.
Jack Gold, principal analyst at J. Gold Associates, named the root cause without softening it for the outlet's readers. "They have a bias to doing everything in house," Gold said, describing Apple's instinct to build a rival to Gemini and GPT-series models from scratch rather than license capability the way it eventually did. That instinct is the same one that kept Apple building its own maps, its own search infrastructure, and its own chips — a strategy that worked when the underlying technology matured on a hardware timeline. Frontier language models did not mature on that timeline, and Apple's typical patience became the two-year gap between promise and delivery.
iOS 26 adoption data tells a quieter version of the same story. Apple's own telemetry, reported through Daring Fireball's tracking of Apple's adoption figures, showed 74% of eligible iPhones running iOS 26 within five months of release — a hair behind iOS 18's 76% pace at the same point a year earlier. The loudest AI marketing push in Apple's history did not accelerate the upgrade cycle at all. It ran essentially flat against a year with no comparable AI story to tell.
The CEO Change That Wasn't Really About Retirement
Tim Cook's exit, announced through an Apple Newsroom release on April 20, 2026, takes effect September 1 — the same month iOS 27 and the rebuilt Siri finally reach the public. Cook becomes executive chairman. John Ternus, senior vice president of hardware engineering for the past five years and a 25-year veteran of the company, becomes Apple's eighth chief executive.
Ternus built his internal reputation on iPad, AirPods, and the Vision Pro's hardware program, not on machine learning research, and that choice reads as a deliberate signal rather than an oversight. Timothy Hubbard, an assistant professor of management at Notre Dame, told CNBC that promoting a hardware leader over software or AI executives suggests Apple's board believes the next phase of AI competition runs through tightly integrated devices rather than the models themselves. Jim McGregor of Tirias Research put the mandate more bluntly in the same reporting: Apple, he said, cannot afford to arrive late again the way it did when ChatGPT reset the industry in 2022.
The board passed over Craig Federighi, Apple's software engineering chief and the executive most publicly associated with the company's AI missteps, along with former COO Jeff Williams, who retired earlier in 2026. Choosing Ternus over both signals something specific: Apple's next CEO is not being asked to build a model that beats Gemini. He is being asked to make sure the devices running Gemini stay indispensable enough that the question never comes up.
What Happens to ChatGPT Inside the New Siri
ChatGPT held the default supplementary role inside Siri since late 2024, handling the complex queries Apple's own models could not manage. Gemini now occupies that position instead, and OpenAI's tool shifts to an optional plugin a user has to invoke rather than the system's automatic fallback, according to reporting on the iOS 26 and iOS 27 rollout timeline.
iOS 27's architecture reportedly opens a path for other model providers to publish first-class integrations rather than sitting behind Siri as a secondary option, which would let OpenAI build a native ChatGPT experience inside Apple's ecosystem on more equal footing with Gemini. Apple has not confirmed this publicly, and the distinction between "optional plugin" and "first-class integration" is still being worked out inside developer documentation that has not fully shipped. Whether that architecture materializes as described, or gets narrowed before iOS 27's public release, remains one of the open variables in this rollout.
Where Apple Intelligence Still Has an Actual Edge
Licensing the reasoning layer does not erase what Apple actually controls, and the privacy architecture underneath the Gemini arrangement is the clearest example. Requests never leave Apple-controlled infrastructure, Google never receives raw queries, and the entire routing decision happens invisibly to the user — a level of data discipline no Android device running Gemini natively can currently match, simply because Google is both the model provider and the data controller on its own hardware.
Distribution is the second edge, and it is larger than any single feature comparison. Apple Intelligence reaches roughly 1.56 billion active iPhones the moment it ships, activated through an update users already trust enough to install at a 74% clip within months. Gemini's own standalone app has to be downloaded, opened, and chosen. Apple's version arrives pre-installed, defaulted, and wrapped inside a decade of accumulated user trust in how the company handles personal data — the exact asset a licensing deal cannot buy and a rival cannot replicate by shipping a better model alone.
Boeing does not build its own jet engines. It buys them from Rolls-Royce and GE, then wins on airframe design, safety record, and the trust airlines place in the finished aircraft. Apple's Gemini bet follows the identical logic: license the hardest component from whoever builds it best, and compete on everything wrapped around it instead.
Whether that comparison holds depends on something Boeing never had to worry about: whether the engine supplier is also building its own competing airframe. Google runs Gemini natively on Android, sells Pixel hardware, and has every incentive to make its own devices the best possible showcase for its own model — the one advantage Rolls-Royce never had over Boeing.
Who This Is For
An iPhone 15 Pro or newer owner deciding whether to update through iOS 27 this September gets a direct answer here: yes, and the reasoning quality jump will be real, because it is coming from a materially larger model than Apple shipped before. A developer building on Apple's platforms needs to watch whether the multi-model plugin architecture actually opens by general availability, since that determines whether a ChatGPT-based product can compete on equal footing inside Siri. An investor weighing Apple's AI position against Google, Microsoft, and OpenAI needs the distribution argument above more than any benchmark score, since Apple is not trying to win on model quality at all.
Google's own AI division is fighting a separate, larger battle over spending and market share — see how Gemini's benchmark lead fits into the broader trillion-dollar contest among OpenAI, Anthropic, and Google.
Verdict
Verdict
Apple is neither catching up nor quietly leading — it has redefined what winning means for itself. Owning the model layer stopped being the goal the moment Apple signed a deal acknowledging it could not build one at Gemini's level on its own timeline. Owning the interface, the privacy guarantees, and 1.56 billion trusting devices became the actual strategy, and that bet is genuinely defensible if Google never turns its engine-supplier position into a reason to undercut Apple's hardware directly. Update this position once iOS 27 ships in September and the reasoning tier faces real-world load for the first time.
Ternus takes over a company that just admitted, through the clearest signal available to a corporation that rarely admits anything, that its next chief executive does not need to understand frontier model training half as well as he needs to understand why 1.56 billion people keep trusting Apple with their data anyway.
What is Apple Intelligence in 2026?
Apple Intelligence is Apple's on-device and cloud AI system spanning iPhone, iPad, Mac, Apple Watch, and Vision Pro. In 2026 it routes requests across three tiers: on-device Apple models for simple tasks, Private Cloud Compute for moderate tasks, and a custom Google Gemini model for complex reasoning inside the rebuilt Siri.
Is Siri now powered by Google Gemini?
Yes. Apple and Google jointly confirmed in January 2026 that a custom 1.2-trillion-parameter Gemini model handles Siri's complex reasoning and planning tasks, while simple requests still run entirely on Apple's own on-device models.
Does Google see my data through Siri?
No raw user data passes to Google under the arrangement. Gemini's reasoning runs through Apple's Private Cloud Compute on Apple-controlled servers, and Apple has said queries are not stored or used to train Google's models.
Which iPhones support Apple Intelligence?
Apple Intelligence requires an A17 Pro chip or newer, which covers iPhone 15 Pro, iPhone 16, and iPhone 17 models. Older iPhones can reach limited cloud-based features but not on-device processing.
Why was Siri's AI overhaul delayed for two years?
Apple promised a more personal Siri alongside iOS 18 in 2024, then pushed the release through iOS 26.4 before folding the full rebuild into iOS 27 in September 2026, after its own models fell short of what it had already promised publicly.
Who is Apple's new CEO and why does it matter for AI?
John Ternus, Apple's hardware engineering chief for five years, becomes CEO on September 1, 2026, replacing Tim Cook. Analysts including Jim McGregor of Tirias Research say Ternus inherits the role specifically to fix Apple's late arrival in the AI race.
Is Apple Intelligence better than Google's own Gemini app?
Not on raw model capability, since both now draw on the same Gemini family. Apple's advantage is privacy architecture and integration across 1.56 billion active devices; Gemini's advantage is reach across platforms beyond Apple hardware.
Can I still use ChatGPT inside Siri?
Yes, but it now sits behind Gemini as an optional plugin rather than the default assistant. Apple's iOS 27 architecture is expected to let OpenAI publish a more fully integrated ChatGPT option for users who prefer it.
Peak of Trending is tracking Apple's AI transition through the iOS 27 launch and the Ternus succession — follow along for what changes next.