Apple has spent eighteen months insisting Siri would eventually catch up to ChatGPT and Gemini. As of September 10, 2026, the company’s own explanation of how it actually caught up is more complicated than “Apple bought Gemini.” According to reporting from MacRumors, 9to5Mac, and other outlets covering Apple’s June 8, 2026 architecture reveal at WWDC, the new Siri is built on five Apple Foundation Models, four of which were trained using outputs from Google’s Gemini frontier models. None of those five models is Gemini itself, and none of them ships as a Google app on the iPhone.
That distinction, small as it sounds, is the whole story. Apple and Google confirmed a multi-year partnership on January 12, 2026, and outlets including TechCrunch reported the deal is worth roughly $1 billion a year to Google. What that deal actually produced, unveiled five months later, is a family of custom Apple models distilled from Gemini rather than a rebadged Gemini Assistant running under the Siri icon. Here is what Apple has confirmed, what Google has confirmed, and what remains attributed to press and analyst reporting.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Apple actually announced about Siri and Gemini
Apple’s newsroom framed the update around “the next generation of Apple Foundation Models, custom-built in collaboration with Google and its Gemini models,” running on-device and through Apple’s Private Cloud Compute servers. That framing matters because it puts Apple’s own infrastructure, not Google’s cloud, at the center of the sentence. The word “collaboration” is doing a lot of work: Apple is not licensing the Gemini app, and it is not routing Siri queries through Google’s consumer AI product. It licensed access to Gemini’s underlying model family to train and refine its own models, the same models Apple began surfacing on the iPhone 18 Pro’s A20 Pro silicon earlier this year.
Google and Apple published a joint statement on the collaboration, describing a multi-year, non-exclusive partnership under which “the next generation of Apple Foundation Models will be based on Google’s Gemini models and cloud technology,” according to Google’s own company blog. Google reiterated that Apple Intelligence would keep running on Apple devices and through Private Cloud Compute, which is Apple’s shorthand for cloud processing that discards user data after each request and never persists it for Google, Apple, or anyone else to retrieve later.
Apple’s own quote on the decision, reported by CNBC, was blunt about the tradeoff: “After careful evaluation, we determined that Google’s technology provides the most capable foundation for Apple Foundation Models and we’re excited about the innovative new experiences it will unlock for our users.” Apple did not say it had built a better model in-house. It said Google’s technology was the best foundation available, and it built on top of that foundation instead of starting from zero.
Five models, not one: how the Siri AI stack actually splits up
The architectural detail that outlets converged on after the June 8 reveal is a five-model system, not a single “Siri brain.” Per 9to5Mac’s breakdown of Apple’s own documentation, Apple built five foundation models in collaboration with Google. Four of them are customized Gemini-derived models optimized to run on Apple Silicon, and the fifth is a larger cloud model that runs on Google-operated servers under Apple’s privacy rules rather than Google’s standard cloud stack.
The routing logic follows a familiar three-tier pattern that Apple has used since the original Apple Intelligence launch in 2025, just with more capable models slotted into each tier. Two of the five models run entirely on-device, handling requests that never need to leave the phone. Two more run on Apple’s Private Cloud Compute servers, Apple’s own confidential-computing infrastructure, for moderately complex requests that need extra horsepower but not frontier-scale reasoning. The fifth and largest model, which Apple has referred to internally as AFM 3 Cloud Pro according to press summaries of the WWDC briefing, runs on Google-dedicated servers using the same Private Cloud Compute privacy guarantees, reserved for the hardest reasoning and world-knowledge queries.
Trained on Gemini, not running Gemini
The technical distinction Apple is pressing hardest is the difference between a model trained using another company’s outputs and a model that simply is the other company’s product with new branding. Analysis cited by 9to5Mac describes the four Apple Silicon-optimized models as “trained using proprietary data with reinforcement learning and refined using outputs from Gemini frontier models.” That is a real and meaningful distinction in machine learning terms: it is closer to distillation, where a smaller model learns to approximate a larger teacher model’s behavior, than to deployment, where a company simply calls someone else’s API at runtime.
Apple’s Craig Federighi, the company’s senior vice president of software engineering, addressed the confusion directly in comments reported by 9to5Mac. “Of course, we don’t have the Gemini app as our app,” Federighi said, drawing a line between what ships on the iPhone and what Google offers as a standalone consumer product. He described the underlying models more plainly elsewhere in the same briefing: “These are the models that are the product of our collaboration with Google.” That is about as precise as Apple has been willing to get in public: collaboration, not deployment.
What is confirmed versus what is still reported and rumored
Because Apple has not published a routing diagram or a technical whitepaper on the five-model system, a meaningful share of the detail circulating since June 8 comes from press reconstruction rather than official documentation. Readers trying to separate fact from informed guesswork should keep the following split in mind.
| Claim | Status | Source |
|---|---|---|
| Apple Foundation Models are custom-built in collaboration with Google and Gemini | Confirmed (Apple newsroom) | Apple, WWDC 2026 |
| Siri and Apple Intelligence continue to run on-device and via Private Cloud Compute | Confirmed (Apple and Google statements) | Apple / Google joint statement |
| Apple is extending Private Cloud Compute to run on Google Cloud infrastructure with Nvidia hardware | Confirmed (Apple security research post) | Apple Security Research blog |
| Partnership is multi-year and non-exclusive | Confirmed (joint statement framing) | Apple / Google |
| Deal is worth roughly $1 billion per year to Google | Reported, not officially priced by either company | Press and analyst reporting |
| Custom Gemini-derived model uses roughly 1.2 trillion parameters | Reported, not confirmed by Apple or Google | Infrastructure and analyst reporting |
| Full Siri rollout timing tied to a specific iOS point release | Reported and speculative; Apple has not named a minor version | Press reporting |
| Apple is developing its own “Baltra” AI server chips as a longer-term Gemini alternative | Reported by supply-chain sources; not addressed in Apple’s Siri statements | Infrastructure reporting |
That table matters more than most of the technical detail in this story, because the gap between “Apple confirmed X” and “press reported X” is exactly where confusion about Siri and Gemini keeps regenerating itself every few weeks. Apple has been consistent about the architecture-level claim (custom models, trained with Gemini’s help, running on Apple’s own infrastructure) and consistently silent on the commercial and hardware-roadmap specifics that would let outsiders independently verify the deal’s scale.
Private Cloud Compute: the privacy argument Apple is leaning on
Private Cloud Compute has been Apple’s answer to the “if it’s in the cloud, is it really private” question since the original Apple Intelligence launch, and it is doing even more work now that a Google-trained model sits inside the stack. Apple’s Security Research team confirmed in a June 8, 2026 post that the company is collaborating with Google and Nvidia to run new Apple Intelligence workloads on Google Cloud, extending Private Cloud Compute’s privacy guarantees to hardware Apple does not physically own or operate. The pitch is that the confidential-computing protections travel with the workload rather than staying tied to Apple-owned data centers.
Outlets describing the flow in more granular terms report that Siri requests are pre-processed on-device, with personal identifiers stripped before anything leaves the phone, and that a de-identified query is what actually reaches the largest, Google-hosted model tier. Apple’s own documentation confirms the privacy architecture and the confidential-compute framing; the step-by-step data-flow diagrams circulating in explainer pieces are journalists’ reconstructions of that architecture rather than verbatim text from Apple. The practical upshot both sides agree on: Apple says Google does not receive names, Apple IDs, contacts, or other persistently linked identifiers as part of this pipeline.
It is worth pointing out what Apple’s Siri AI does not draw on, per 9to5Mac’s reporting: the new models rely on Apple’s own knowledge graph for world-knowledge queries, not Google Search or Google’s knowledge graph. That is a second, quieter separation between “Siri uses Gemini-derived models” and “Siri uses Google’s search index,” and it is one Apple has been careful to keep distinct in briefings.
Why Apple picked Gemini over building everything in-house
The honest answer, based on Apple’s own public statement, is that Apple concluded it could not close the reasoning gap with OpenAI and Google on its own timeline. Siri’s 2024 and 2025 stumbles, including the well-documented delay of the more personalized Siri originally promised for iOS 18, left Apple visibly behind Gemini and GPT-5-class models on multi-step reasoning benchmarks. Rather than keep iterating on a from-scratch Apple model and risk another multi-year delay, Apple licensed access to a more capable foundation and put its own training, privacy, and on-device optimization work on top of it.
For Google, the arrangement is a distribution win that is hard to replicate through any other channel: Gemini-derived intelligence now sits inside thet software running on every iPhone capable of Apple Intelligence, without Google needing to win users away from Siri as a habit. Press coverage of the January 2026 announcement framed it as one of the most consequential distribution deals in the AI market since Google itself paid Apple to remain the default search engine in Safari, a deal that has drawn its own antitrust scrutiny for years.
Where OpenAI’s ChatGPT integration fits now
Apple’s 2025 Apple Intelligence launch leaned on OpenAI’s ChatGPT as an optional, opt-in extension: when Siri could not answer something confidently on its own, users could choose to have the request handed off to ChatGPT, with a visible prompt each time. That relationship has not been terminated, but the center of gravity has clearly shifted. The 2026 architecture positions Gemini-derived Apple Foundation Models as the default reasoning layer for Siri and system-wide intelligence features, while ChatGPT remains available as a secondary, user-invoked option rather than the engine underneath Siri itself.
That is a meaningful reversal from where things stood a year earlier, when OpenAI looked like Apple’s primary external AI partner. Analysts covering the shift have described it as Apple choosing a deeper infrastructure and model-training relationship with Google over a shallower, API-style integration with OpenAI, largely because the Google deal gave Apple training data access and model customization rights that a simple ChatGPT handoff never offered. That said, OpenAI’s own GPT-6 Astra launch this year kept the ChatGPT option relevant enough that Apple had little reason to drop it entirely.
Competitive comparison: Siri’s new stack versus rival assistants
Putting the Apple, Google, Samsung, and Amazon assistant strategies side by side shows just how unusual Apple’s approach is. Most competitors either built their own frontier model in-house or fully embedded a partner’s product under its own name. Apple is doing neither.
| Assistant | Underlying model approach | On-device processing | Primary cloud partner |
|---|---|---|---|
| Siri (Apple Foundation Models, 2026) | Custom models distilled from Gemini, trained by Apple | Yes, two of five models run fully on-device | Google Cloud (Private Cloud Compute extension) |
| Google Assistant / Gemini on Android | Native Gemini models, Google’s own product | Limited, mostly cloud-routed | Google Cloud (native) |
| Amazon Alexa+ | Built on Anthropic’s Claude models under a commercial license | Limited, primarily cloud-routed | AWS, with Anthropic model licensing |
| Samsung Bixby / Galaxy AI | Mix of in-house models and licensed Gemini features | Partial, varies by feature | Google Cloud (partnership) |
| Microsoft Copilot | Built on OpenAI’s GPT models under a commercial license | Minimal, cloud-first | Microsoft Azure, with OpenAI model licensing |
The pattern across the industry is that almost nobody is training a truly from-scratch frontier model anymore for consumer assistants; the economics of doing so don’t pencil out against licensing a leading lab’s technology and specializing it. Google’s own Gemini 3.8 Flash launch earlier this year showed how aggressively the company is pricing access to that same model family for third parties, which makes Apple’s decision to license rather than build look even more like the economically rational choice. Apple’s version of that licensing model is more restrictive and more privacy-engineered than most, which tracks with the company’s decade-long marketing position on user data, but the fundamental strategy of “license a frontier model, then customize” now looks like the industry default rather than the exception.
Historical context: Siri’s long road to this point
Siri launched in 2011 as one of the first mainstream voice assistants, and for most of the following decade it was widely regarded as having fallen behind Google Assistant and Amazon Alexa on natural-language understanding. Apple’s 2024 WWDC promised a more personalized, context-aware Siri as part of the original Apple Intelligence rollout, tied to iOS 18. That feature slipped repeatedly through 2024 and 2025, becoming one of the most visible product delays in recent Apple history and drawing direct criticism from developers and reviewers who had expected the capability at launch.
The January 2026 Google partnership announcement effectively closed that chapter by acknowledging, without saying so explicitly, that Apple’s in-house model work had not caught up to what Gemini and GPT-5-class systems could already do. The June 2026 WWDC reveal of the five-model Apple Foundation Models architecture was Apple’s attempt to show that the partnership produced something distinctly Apple rather than a rebranded Google product, a distinction the company has repeated in nearly every public comment since, including at Apple’s September 9 hardware event, where executives kept the AI messaging centered on Apple Intelligence branding rather than Gemini.
Market and investor reaction
Financial coverage of the January 12, 2026 announcement described the market reaction as incrementally positive for both companies, though for different reasons. Alphabet’s stock drew attention from investors who saw the deal as validation that Gemini’s underlying technology was strong enough for a rival of Apple’s stature to license at scale, reinforcing Google’s position in the broader AI infrastructure market alongside its cloud and advertising businesses. Apple’s stock reaction was more muted, consistent with a company that was closing a competitive gap rather than announcing new growth, but analysts broadly characterized the move as removing a lingering overhang on Apple Intelligence’s credibility heading into the second half of 2026, a credibility boost that showed up later when Apple’s Mac revenue hit a record quarter partly on demand from AI labs buying Apple hardware.
Neither company has published exact stock-price movement figures tied specifically to the Siri-Gemini news, and any percentage swings reported in the days following the announcement reflect broader market conditions as much as the deal itself, so treat any single-day AAPL or GOOGL percentage figure circulating online with caution unless it comes from an official market data provider.
What developers and enterprise IT teams should actually do with this
For engineering teams building on top of Apple’s platforms, the practical takeaway is narrower than the headlines suggest. Apple has not opened direct API access to the Gemini-derived Foundation Models the way OpenAI or Google expose their own frontier models; Apple’s on-device Foundation Models framework, introduced alongside Apple Intelligence, remains the sanctioned way for third-party apps to tap into Siri-adjacent intelligence, and that framework’s capabilities are what actually ship to developers, not the internal training lineage of the models behind it.
Enterprise security and compliance teams evaluating Apple Intelligence for regulated environments should focus on the Private Cloud Compute privacy claims Apple has published and independently audited rather than on the Gemini training relationship, since the data-handling guarantees are what actually determine regulatory exposure. Apple has published technical detail on Private Cloud Compute’s architecture and has opened parts of it to external security researchers in the past, which is the more relevant due-diligence trail for compliance purposes than press speculation about parameter counts.
Predictions: where the Apple-Google AI relationship goes next
- Apple will continue drawing a hard public line between “trained with Gemini’s help” and “running Gemini,” and will keep correcting outlets and users who conflate the two, since brand control over Siri’s identity is a stated priority for the company.
- Expect Apple to expand the on-device share of the five-model architecture over time as its own Apple Silicon Neural Engine capacity grows, gradually shrinking reliance on the Google-hosted cloud tier for anything short of the hardest reasoning tasks.
- The reported in-house “Baltra” AI server chip effort, if it materializes on the timeline described in supply-chain reporting, is likely to be framed publicly as complementary capacity rather than a replacement for the Google partnership, at least through the current multi-year deal term.
- OpenAI’s ChatGPT integration will likely persist as an opt-in feature rather than disappear, since Apple has an incentive to keep multiple frontier-model relationships live for negotiating leverage and redundancy.
- Regulatory scrutiny of the Apple-Google relationship is likely to grow, given that antitrust regulators in the US and EU have already scrutinized the separate Google-pays-Apple-for-Safari-default arrangement for years; a second major financial relationship between the same two companies is a natural next target for scrutiny.
The bottom line on Siri and Gemini
Does Siri use Google Gemini? The most accurate short answer, based on what Apple and Google have actually confirmed rather than what press coverage has extrapolated, is: Siri’s newest AI capabilities are built on five Apple Foundation Models, four of which were trained and refined using outputs from Google’s Gemini frontier models, and the fifth of which runs on Google-operated infrastructure under Apple’s own privacy architecture. None of the five is the Gemini app or Gemini Assistant running under a different name. Apple picked Google’s technology as the strongest available foundation, licensed access to it in a multi-year, non-exclusive deal reported at roughly $1 billion a year, and then spent the better part of 2026 building something it can plausibly call its own on top of it. Whether that distinction matters to the average iPhone user asking Siri to summarize an email is a separate question from whether it’s technically accurate, and on the technical question, Apple’s own documentation and executives have been consistent since January. For more on how the leading AI models stack up against each other heading into the back half of 2026, see our AI model comparison hub.
Frequently asked questions
Does Siri actually run on Google Gemini?
No. Siri runs on Apple Foundation Models, a family of five custom models. Four were trained using outputs from Google’s Gemini frontier models and optimized to run on Apple Silicon; the fifth is a larger model hosted on Google-operated servers under Apple’s Private Cloud Compute privacy rules. None of the five is the Gemini app or Gemini Assistant itself.
Is the Gemini app installed on the iPhone as part of this deal?
No. Apple’s Craig Federighi directly addressed this point, saying Apple does not have “the Gemini app as our app.” The partnership covers model training and licensing, not shipping Google’s consumer Gemini product on iOS.
How much is Apple paying Google for this partnership?
Press and analyst reports place the figure at roughly $1 billion per year, but neither Apple nor Google has published an official dollar figure. Treat that number as reported, not confirmed.
Does Google get access to Siri user data through this partnership?
Apple says no. Requests are processed through Apple’s Private Cloud Compute architecture, which Apple describes as stripping personal identifiers before any data reaches the Google-hosted model tier and not retaining data after a request completes. Apple has extended these privacy guarantees to the Google Cloud infrastructure it uses for the largest model tier.
What happened to Apple’s ChatGPT integration with Siri?
It is still available as an optional, user-invoked feature, but it is no longer the primary way Siri accesses external AI reasoning. The Gemini-derived Apple Foundation Models are now the default intelligence layer, with ChatGPT positioned as a secondary option.
When did Apple and Google confirm this partnership?
Apple and Google issued a joint statement on January 12, 2026 confirming a multi-year, non-exclusive collaboration. Apple detailed the resulting five-model Apple Foundation Models architecture publicly at WWDC on June 8, 2026.
Can third-party developers access the Gemini-derived Siri models directly?
Not directly. Apple’s sanctioned path for developers remains its on-device Foundation Models framework introduced with Apple Intelligence, which exposes Apple’s own model capabilities to apps rather than granting direct access to the internal Gemini-derived training lineage.
Is Apple still building its own AI models from scratch?
Supply-chain and infrastructure reporting describes an in-house AI server chip effort, sometimes referred to as “Baltra,” that could reduce Apple’s reliance on the Google partnership over time. Apple has not addressed this reporting directly in its official Siri or Apple Intelligence communications.
Related Coverage
- Mercury 2.5 vs Gemini Diffusion: LLMs Hit 10x Speed [2026]
- ChatGPT vs Copilot vs Gemini: 30-Store Checkout Gap [2026]
- GPT-6 Astra vs Claude Opus 5 vs Gemini 3.8 Flash: 107x Gap [2026]
- Google Pics Rivals Canva Weeks After $10B Valuation Cut [2026]
- GPT-5.6 Sol vs Fable 5.1 vs Gemini 3.8 Flash: 30-Point Gap [2026]
