At today’s WWDC 2026 keynote, Apple confirmed what Bloomberg’s Mark Gurman had been reporting for three months: a custom 1.2-trillion-parameter Google Gemini model is now the default brain behind Siri on roughly 1.4 billion active iPhones. The same keynote also confirmed the part Gurman didn’t fully spell out — iOS 27 ships with a system-level Extensions framework that lets the user, or an MDM administrator, swap Gemini out for ChatGPT or Claude across every Siri surface.

This is two announcements stacked into one slide. A default model change at platform scale. And a choice layer that turns the default into a procurement decision.

For consumer press, the headline is Gemini-on-iPhone. For anyone running a corporate fleet, the headline is buried two layers down: Siri 2.0 is a full chatbot — persistent memory, on-screen awareness, cross-app actions, web search, file analysis — running by default on every managed iPhone in your organization. Whether your existing AI usage policy mentions Siri or not, it does now.

Quick Summary: What WWDC 2026 Actually Changed

Detail	Info
Date	June 8, 2026 (WWDC 2026 keynote)
Default Siri model	Custom 1.2T-parameter Google Gemini (replaces Apple Foundation Models for upstream queries)
Reported license cost	~$1 billion per year (per Bloomberg)
Extensions framework	System-level setting — choose ChatGPT, Gemini, or Claude as the Siri model
Data boundary	All third-party model inference runs inside Apple Private Cloud Compute — stateless, ephemeral, no training rights
Siri 2.0 capabilities	Persistent memory, on-screen awareness, cross-app actions, web search, file analysis
Public release	iOS 27 ships September 2026
Install base affected	~1.4 billion active iPhones

Bottom line: Apple just shipped a chatbot layer to every iPhone by default and handed enterprise IT a vendor choice they didn’t ask for and now have to make.

---

What Apple Announced at WWDC 2026

The keynote story has two threads, and the consumer press is going to braid them in a way that buries the part that matters for AI tool buyers. Let’s keep them separate.

Thread one: Gemini is the new default. Apple’s Foundation Models (the on-device models that handle small Siri tasks, dictation, Writing Tools, and the like) stay in place. What changed is the upstream model. When Siri hits a query that needs frontier-scale reasoning, world knowledge, or longer-context processing, that request used to fall back to Apple’s own server-side models (with the Gemini-powered Siri overhaul Apple shipped in iOS 26.4). In iOS 27, the default upstream is a custom Gemini variant Apple licensed from Google. Bloomberg’s reporting put the license at roughly $1 billion per year, and Apple’s keynote did not contradict the number.

The model is reportedly a 1.2-trillion-parameter custom variant — built by Google for Apple’s privacy and latency requirements, not a stock Gemini SKU. The architectural detail Apple emphasized in the keynote was the data envelope: all inference runs inside Apple’s Private Cloud Compute on Apple Silicon servers. Per Apple, queries are stateless, ephemeral, and unavailable to Google for any training or product purpose. The contract terms — Apple’s word, not Google’s — forbid using Apple user queries as training data for future Gemini models.

That’s the trust architecture Apple is asking enterprises to accept. The model is Google’s. The data boundary is Apple’s. Google sees the query, computes the answer, returns it, and the system holds no persistent record. If you believe Apple’s Private Cloud Compute attestation chain — and the security research community has largely treated it as credible — the architecture is materially different from sending a query to consumer Gemini and trusting Google’s data policies.

Thread two: Extensions makes it a choice. This is the part the consumer angle obscures. iOS 27 ships with a system-level setting called Extensions that lets the user replace the default Gemini model with ChatGPT or Claude across every Siri surface. The architecture is the same regardless of which provider is selected: third-party model, Private Cloud Compute boundary, no training on user data. The difference is which vendor’s API is on the other side of the boundary.

So Apple’s existing OpenAI deal, the headline of WWDC 2024, is now one option among three. ChatGPT is no longer the privileged upstream. It’s a checkbox.

Why This Is Procurement, Not Product

The framing the consumer press will run with is “iPhone users get to pick their AI.” That framing is true and almost completely misses the point.

Here’s what’s actually new. Before WWDC 2026, Siri was a feature. It had real limitations, it occasionally embarrassed Apple, and it lived in a small corner of the iOS experience. Most people had stopped using it for anything beyond timers and music. A corporate IT team writing an AI usage policy could reasonably ignore Siri. The model behind it was Apple’s own, the surface area was narrow, and the data flows were inside the Apple boundary.

After WWDC 2026, Siri is a full chatbot. Persistent memory. On-screen awareness across every app. Cross-app actions. Web search. File analysis. That’s not “Siri got better at timers.” That’s ChatGPT-equivalent capability, sitting one button-press away on every iPhone in your fleet, by default, with a Google model on the back end.

This converts mobile AI governance from optional to mandatory. Three reasons.

Every employee is now an AI user. Whether or not your company has approved any AI tools, whether or not employees have ChatGPT or Claude accounts, every employee with an iPhone is using a frontier-grade chatbot. Pressing the side button counts as a tool invocation. The Writing Tools menu counts. The “summarize this PDF” Siri shortcut counts. The corporate AI usage policy that lists which tools are sanctioned now has to either ban Siri (which is functionally banning the iPhone) or sanction Siri (which means sanctioning the upstream model).

The data flow goes outside your perimeter. Even with Apple’s Private Cloud Compute envelope, a Siri query containing sensitive customer data — an attached PDF, a screenshot, a long voice memo about a deal — leaves the device, hits Apple’s servers, gets passed to the selected model provider’s compute layer, and returns. The architecture is privacy-preserving by Apple’s design. It is also, factually, a data flow that didn’t exist in your threat model six months ago. Your DLP rules, your data residency requirements, your contractual obligations around customer data — all of that has to reckon with Siri’s expanded scope.

The vendor choice is a real choice. Extensions makes the model selection explicit, which means somebody has to make it. If you’re an enterprise with managed iPhones, the question “what AI model is allowed to process company data through Siri?” now has three answers, and your security and procurement team owns picking one. That’s a vendor decision, not a Siri feature setting. It interacts with your existing AI vendor contracts, your data processing agreements, and whatever sectoral compliance rules apply to your industry.

The Three-Way Vendor Choice in Detail

Each Extensions option lands differently for an enterprise buyer.

Google Gemini (default)

The 1.2T-parameter custom Gemini variant is the default. For consumer iPhones in the U.S., this is what’s running unless the user changes it. For enterprise iPhones, this is what’s running unless IT changes it.

Practical implications. Most enterprises don’t currently have a Google Gemini enterprise contract, even if they use Google Workspace. The default selection makes Google a data processor for every Siri query routed through Extensions, mediated through Apple’s Private Cloud Compute. The contract Apple negotiated reportedly prohibits training on Apple user queries — but that’s Apple’s contract with Google, not your contract with Google. Your existing vendor management framework probably doesn’t have a row for “Google as a sub-processor of Apple as a data processor of our employees’ Siri queries.” Now it does.

For organizations already deep on Google Workspace and Vertex AI, this is the least friction option — the vendor relationship exists, the data processing agreements probably cover it (or can be amended to cover it), and the procurement work is incremental.

OpenAI ChatGPT

Apple’s original OpenAI integration shipped in iOS 18 in 2024 and held the privileged Siri upstream slot until iOS 26.4, when Gemini replaced it as the default upstream model in spring 2026. iOS 27 Extensions formalizes that shift into an explicit user and MDM toggle — ChatGPT is now one of three named options, not the assumed upstream. Switching to ChatGPT through Extensions restores roughly the experience iPhone users had before the iOS 26.4 overhaul, with the Siri 2.0 architecture wrapped around it.

For enterprises with existing ChatGPT Enterprise contracts, this is the path of least surprise. The vendor relationship is in place, the data processing terms are negotiated, and the model behavior is the one your AI usage policy was probably already written around. The trade-off is what OpenAI’s roadmap looks like under the billing structure that shifted in May and how the Anthropic IPO and capacity dynamics reshape the competitive picture between now and September.

Anthropic Claude

Claude as a Siri upstream is the genuinely new option. Anthropic has been positioning Claude as the enterprise-trust frontier model — slower-paced product cadence, more conservative safety posture, stronger long-context behavior than the alternatives. iOS 27 Extensions is the first time Claude is one toggle away from being the default AI on a 1.4-billion-device platform.

For enterprises that already standardized on Claude — and there are more than there were a year ago, particularly in legal, financial services, and regulated healthcare — this is the option that aligns Siri with the vendor you’ve already chosen. The procurement question becomes “does our Claude enterprise agreement extend to Siri-routed queries through Apple Private Cloud Compute?” That’s a contract conversation worth having before September.

What Siri 2.0 Actually Does

The capability set Apple demoed is what makes this an enterprise question instead of a consumer feature update. The headline capabilities:

1. Persistent memory. Siri now remembers context across sessions — projects, preferences, ongoing tasks, recurring contacts. Same architecture pattern as the ChatGPT memory rewrite that shipped last week, with Apple’s data boundary wrapped around it.
2. On-screen awareness. Siri can see and act on whatever is on the screen. The capability shipped tentatively in iOS 26.4; iOS 27 expands it to third-party apps via a published SDK.
3. Cross-app actions. Siri can chain operations across multiple apps without the user re-stating context. Compose a draft in Mail, schedule the follow-up in Calendar, log the contact in CRM, all from one voice command.
4. Web search. Siri can run live web queries through the selected upstream model and incorporate the results into its response. Citation behavior depends on which model is selected.
5. File analysis. Siri can ingest documents, images, audio files, and PDFs from any iOS file picker and reason over them in context. This is where the data perimeter conversation gets most acute.

The composite is a full chatbot — equivalent in scope to a Claude or ChatGPT desktop client — running on every iPhone, with the model identity selectable. That’s the device fleet that walks through your office’s front door every morning. Half of it is BYOD. The other half is on your MDM enrollment.

The MDM and Compliance Question

Apple did confirm at the keynote that Extensions is exposed to MDM. IT administrators can enforce a specific model selection across an enrolled device fleet, lock the choice so users can’t override it, or disable third-party Extensions entirely and pin to Apple’s Foundation Models for the upstream tier. The MDM controls land in the iOS 27 developer beta as of today and will be in production for Microsoft Intune, Jamf, and the rest of the EMM ecosystem by the September public release.

What MDM does not do: protect employees’ personal iPhones. The same employee whose corporate iPhone is pinned to Claude with on-device file access disabled is using Gemini by default on their personal phone, with the same set of company-data-adjacent contacts, photos, and messages. The BYOD policy your security team wrote in 2023 is probably not equipped for this.

The compliance angle stacks on top. For US enterprises tracking the Great American AI Act’s discussion draft, Siri is now a deployment surface for frontier AI. For organizations subject to EU AI Act high-risk obligations, the choice of upstream model interacts with the risk categorization of the use case. For organizations with data residency or sector-specific compliance constraints, the question is whether Apple’s Private Cloud Compute architecture satisfies your residency requirements when the underlying model provider’s compute may not be in your jurisdiction.

None of these are unanswerable questions. They’re just questions your AI governance program now has to answer for Siri specifically, on a deadline that aligns with iOS 27’s September release.

What This Says About the Frontier Model Market

WWDC 2026 reframes the competitive picture between frontier labs in a few specific ways.

OpenAI loses default position on Apple. The privileged upstream slot — worth an enormous distribution premium for the last two years — is gone. ChatGPT is now one of three options, not the assumed answer. OpenAI’s enterprise distribution strategy and direct-to-developer channels still exist, but the consumer-app distribution moat that came from being inside Siri shrinks meaningfully.

Google wins default position and pays for it. A $1B/year license is real money even for Google, but the strategic value of being the default model behind every U.S. iPhone is hard to overstate. It also gives Google a real-world dataset of how a frontier model performs across the broadest possible consumer surface — without violating the no-training contract terms, the operational telemetry alone is valuable.

Anthropic gets distribution it didn’t have. Claude’s distribution story has been heavy on direct enterprise deals, partner integrations, and API revenue. iOS 27 Extensions puts Claude one settings toggle away from 1.4 billion users — most of whom would never have downloaded the Claude app. That’s a fundamental change in Anthropic’s consumer accessibility, even with Gemini holding the default. It compounds with the Anthropic competitive positioning we covered around their IPO earlier this month.

The labs that didn’t get in are the news that didn’t happen. xAI’s Grok was widely rumored as the fourth Extensions partner. It’s not in the launch announcement. Whether that’s a delay, a contract issue, or Apple’s editorial choice will be the subplot to watch through the September release. Meta, DeepSeek, and Mistral are absent for the obvious reasons — no consumer trust pitch, no Apple-grade data boundary commitments, no negotiating leverage with Apple.

What to Do This Week

If you’re running an AI program inside an enterprise, three concrete moves.

Open the Extensions conversation with security and procurement before September. The default for unmanaged devices in your fleet will be Gemini. The right answer for your organization may be Gemini, may be Claude, may be ChatGPT, may be “disable Extensions entirely and pin to Apple Foundation Models.” Whichever it is, the decision should be made deliberately on data residency, vendor contract, and compliance grounds — not by accepting the iOS 27 default through inaction.

Update your AI usage policy to name Siri specifically. Most existing AI policies enumerate tools by name and miss OS-level capabilities. iOS 27 forces the rewrite. Decide what classes of data are allowed through Siri-mediated queries, document the answer, and communicate it to employees before September. The policy that doesn’t mention Siri is the policy that gets overrun on day one of the iOS 27 rollout.

Audit your data processing agreements. Apple’s contract with Google does not extend Google’s obligations to your relationship with Google. If Gemini becomes the model behind your employees’ Siri queries — through default or through MDM choice — you need to verify what data processing terms apply. Same exercise for Claude (Anthropic) and ChatGPT (OpenAI). The DPAs you signed in 2024 probably don’t anticipate this surface. Update them before, not after.

Frequently Asked Questions

Is Gemini the only model that powers Siri in iOS 27?

No. Gemini is the default upstream model when Siri needs frontier-scale reasoning. Apple’s own on-device Foundation Models still handle small tasks — quick replies, Writing Tools, on-device dictation. And users (or MDM administrators) can switch the upstream model to ChatGPT or Claude through the Extensions setting.

What does Apple’s Private Cloud Compute actually do?

Private Cloud Compute is Apple’s server-side inference architecture designed to be cryptographically verifiable and stateless. Queries are processed in isolated enclaves on Apple Silicon servers, the model provider sees the query content but has no persistent record, and Apple does not log or retain query content either. The model provider — whether Google, OpenAI, or Anthropic — is bound by Apple’s contract terms, including no training on Apple user queries.

When does iOS 27 ship?

September 2026, alongside the iPhone 18 lineup. Developer betas are available starting today, June 8, 2026. Public beta is expected in July.

Can enterprise IT disable Extensions entirely?

Yes. MDM controls in iOS 27 let administrators disable third-party Extensions, pin the upstream model to a specific provider, or lock the choice so users can’t override it. The same controls land in Intune, Jamf, and other EMM platforms by the September release.

Does Gemini get to train on Apple user queries?

Per Apple’s announcement and the reported contract terms, no. Google receives Apple-routed Siri queries through Apple’s Private Cloud Compute envelope, processes them, and returns answers. The contract prohibits using these queries for model training. The same terms apply to OpenAI and Anthropic when those providers are selected through Extensions.

How does this affect existing ChatGPT integration on iPhones?

The iOS 26.4 ChatGPT integration was a single-vendor handoff. In iOS 27, ChatGPT becomes one of three Extensions options. Users currently using ChatGPT through Siri will keep that experience if they switch ChatGPT to their selected Extension; otherwise the default is Gemini.

What happens to Apple’s own AI development?

Apple’s Foundation Models continue to handle on-device tasks. The frontier-model upstream is where Apple decided licensing was the faster path. Apple’s keynote did not suggest the company is abandoning internal frontier-scale model development, but the operational reality is that Google’s Gemini is the production model behind Siri 2.0 for at least the iOS 27 lifecycle.

Should I delay my company’s AI tool decisions until after WWDC 2026 settles?

No. The iOS 27 Extensions framework is additive to your AI vendor decisions, not a replacement for them. Whatever model you’ve standardized on for direct enterprise use (Claude, ChatGPT, Gemini) probably remains the right answer through Siri Extensions. The new work is making the alignment explicit and updating your policies and contracts to cover the Siri surface specifically.

Our Take

WWDC 2026 is not a Siri update. It’s a procurement event dressed up as a consumer announcement.

The Gemini-as-default story is the part that will dominate consumer coverage, and it’s a real story — a $1B/year licensing deal that quietly admits Apple’s frontier-model ambitions ran behind schedule and that Google’s Gemini is good enough to live behind 1.4 billion iPhones. That’s a meaningful concession. It’s also, increasingly, the right concession. Apple does not need to win the frontier-model race to win the platform race. Owning the data boundary while licensing the model is a defensible strategic position. The Private Cloud Compute architecture is genuinely impressive, and the no-training contract terms address most of the data-flow objections a reasonable enterprise would raise.

The Extensions framework is the part that matters more. Making the upstream model a settings toggle — and exposing that toggle to MDM — converts the AI vendor decision from a developer choice to a platform-level configuration. Every enterprise running iPhones now has a new procurement decision to make, on Apple’s timeline, with Apple’s defaults stacking the deck toward Google. The right enterprise response is to make the choice deliberately on data residency, vendor contract, and compliance grounds, not to accept the default through inaction. The wrong response is to treat Siri 2.0 as a Siri update and ignore the procurement implications.

For the frontier labs, the strategic picture is clearer than it’s been in eighteen months. Gemini wins default distribution and pays handsomely for it. ChatGPT loses its privileged Apple position and has to compete on merit inside Extensions. Claude gets a distribution channel Anthropic could not have built on its own and earns the chance to convert iPhone users it would never have reached otherwise. Whichever model ends up on the most iPhones in 2027 will tell us something real about which lab the market trusts when the choice is presented as a checkbox rather than a brand loyalty.

The September release is when this ships. The procurement conversations should already be in progress. If your AI governance program doesn’t have a Siri row in its vendor matrix by Labor Day, you’re late.

---

Last updated: June 8, 2026. Sources: Apple WWDC 2026 keynote · Apple Private Cloud Compute architecture overview · Bloomberg’s reporting on the Gemini license and Extensions framework · Microsoft Intune · Jamf.

Related reading: iOS 27 Turns AI Tools Into a System Setting · Siri iOS 26.4 Review: Apple’s Gemini-Powered AI Overhaul · Anthropic’s $965B IPO: What It Means for Claude Users · Great American AI Act: What It Means for Tool Buyers · Enterprise AI Deployment 2026 · Anthropic vs OpenAI 2026 · ChatGPT Memory Just Got a Major Overhaul