TL;DR: Apple enters the AI agent arena at WWDC 2026 with a Siri rebuilt on Google Gemini ($1B/year deal), a standalone Siri chatbot app, App Store rule changes for AI agents, and a privacy-first on-device architecture. It’s the company’s most consequential AI moment — and it may redefine how 2.2 billion devices interact with autonomous agents.
Tomorrow morning at 10 a.m. PT, Tim Cook takes the stage at Apple Park for what is almost certainly his last WWDC keynote as CEO. The stakes are higher than any WWDC in memory. After two years of watching OpenAI, Anthropic, and Google ship increasingly capable AI agents while Siri languished as a punchline, Apple is expected to unveil a fundamentally rebuilt AI stack — and early signs suggest it will look nothing like what the competition has built.
The $1 Billion Siri Reboot
The centerpiece is Siri. Not an incremental upgrade. A ground-up rebuild.
On January 12, 2026, Apple and Google issued a joint statement confirming a multi-year partnership: Google’s custom Gemini models would become the foundation for the next generation of Apple Intelligence features, including a completely reimagined Siri. Bloomberg’s Mark Gurman pegged the deal at approximately $1 billion per year — making it one of the largest AI infrastructure deals in history.
Apple reviewed multiple options — including doubling down on its own models and expanding the existing OpenAI partnership — before concluding that Google’s Gemini “offered the most capable foundation.” The irony is thick: the company that built its brand on vertical integration is now licensing its AI brain from its largest platform rival.
What does Gemini-powered Siri actually do? According to leaked Bloomberg mockups and MacRumors reporting, the new Siri arrives in three forms:
- “Search or Ask” overlay: Swipe down from the center of any screen and a glowing pill appears in the Dynamic Island. It expands into a transparent card that shows web results, images, and notes — and can transition into a full conversation mode.
- Standalone Siri app: A dedicated chatbot interface resembling ChatGPT or Claude, with text and voice input, file attachments (PDFs, images, documents), and conversation history with auto-delete options (30 days, 1 year, or forever). Every conversation syncs via iCloud.
- Deep system integration: Siri gains personal context — it can read your emails, messages, calendar, and files to answer questions like “What’s my passport number?” or “Show me the files Eric sent last week.” On-screen awareness lets it act on whatever you’re viewing: add an address from a message to Contacts, send a photo you’re looking at, or draft an email from a conversation you’re reading.
Apple is also building Siri Extensions, a framework that lets users choose third-party AI models — ChatGPT, Claude, Gemini — as alternative backends for Siri requests, Writing Tools, and Image Playground. Each will use a distinct voice so users know which AI is responding. This is a quiet but radical move: Apple is positioning itself not as an AI model provider, but as the runtime layer that every model must pass through to reach users.
Apple Intelligence: The Privacy-First Agent Platform
While Google and OpenAI race toward always-on cloud agents — Google’s Gemini Spark, launched at I/O 2026, runs 24/7 on dedicated cloud VMs — Apple is betting on the opposite architecture.
Apple Intelligence, first introduced at WWDC 2024, is built on a simple principle: process on-device whenever possible, fall back to Private Cloud Compute only when necessary. The next generation of Apple Foundation Models, now rebuilt on Gemini, extends this further. On-device LLMs (internally codenamed “Linwood”) handle the majority of user requests without ever leaving the phone. Complex reasoning tasks are routed to Apple’s PCC servers — ephemeral, verifiable, and cryptographically attested — not to Google’s cloud.
The privacy narrative isn’t just marketing. It’s Apple’s only viable competitive wedge. When Siri reads your emails to answer “what’s my passport number,” it does so on-device. When it analyzes your screen to act on a message, the pixels stay local. This is fundamentally different from ChatGPT or Claude, where every query round-trips through a data center.
But there’s a trade-off: on-device models are smaller, and smaller models are less capable. Apple’s bet is that the Gemini foundation — a 1.2 trillion parameter model, according to Bloomberg — combined with Apple’s on-device fine-tuning, can deliver “good enough” intelligence for 90% of daily tasks while keeping the privacy story intact. The remaining 10% of hard problems go to PCC, and Apple maintains its differentiation.
Whether consumers actually care about this distinction is an open question. ChatGPT’s iOS integration, launched in 2024, has seen reportedly low usage according to The Information — suggesting users haven’t found a compelling reason to use AI inside Apple’s ecosystem rather than as a standalone app.
The App Store Opens to AI Agents
The wildcard at WWDC 2026 may not be Siri at all. It may be the App Store.
On May 13, The Information reported that Apple is actively designing a new framework to allow AI agent apps — including “vibe coding” tools that generate software from natural language — onto the App Store while maintaining its security and privacy standards. This follows a contentious March 2026 decision in which Apple began blocking updates for several popular vibe coding apps, citing rules that prohibit apps from executing code that alters their own functionality.
The timing suggests that Apple recognized it was falling behind. Gartner projects that 40% of enterprise apps will feature task-specific AI agents by the end of 2026. If the App Store cannot host agent-based apps, it risks becoming irrelevant to the next wave of mobile software.
Details on the new framework are scarce, but the strategic implications are enormous:
- Agent distribution: Apple controls the only app distribution channel for 2.2 billion active devices. Opening it to AI agents creates the world’s largest agent marketplace overnight.
- Commission questions: Apple has told some developers it won’t charge commissions “during the early stages” of Siri integration, but “fees are a possibility in the future.” Chinese giants Baidu, Alibaba, and Tencent have already held talks but are reportedly wary of Apple’s commission model.
- Safety guardrails: Apple’s framework aims to prevent the kind of “rogue agent” behavior seen on other platforms — agents that delete emails, overwrite files, or take unauthorized actions. This is the App Store review model applied to agent behavior.
If Apple gets this right, it becomes the gatekeeper for the agent economy. If it gets it wrong — too restrictive, too extractive — developers will route around it, as they have with progressive web apps and browser-based agents.
How Apple’s Approach Compares
| Dimension | Apple | OpenAI | Anthropic | |
|---|---|---|---|---|
| Architecture | On-device first, PCC fallback | Cloud-first | Cloud-first | Hybrid (Spark 24/7 VMs) |
| Model Strategy | Licensed (Gemini) + own fine-tuning | Proprietary (GPT-5.4) | Proprietary (Claude Opus 4.x) | Proprietary (Gemini 3.x) |
| Agent Model | OS-level runtime + Extensions | Codex platform + plugins | MCP protocol + managed agents | Spark ambient + Workspace |
| Distribution | App Store (walled garden) | ChatGPT + API + Codex Sites | API + Claude app + enterprise | Gemini app + Google Workspace |
| Privacy Story | Core differentiator | Improving (ChatGPT biz) | Strong (constitutional AI) | Mixed (cloud-dependent) |
| Revenue Model | Hardware + services + commissions | Subscriptions + API + enterprise | API + enterprise + credits | Ads + cloud + subscriptions |
Apple’s playbook is different from all three competitors in one crucial respect: it doesn’t need to monetize AI directly. Every Siri query that makes an iPhone more useful sells more iPhones. Every agent that runs locally sells more Apple Silicon. Apple can afford to give AI away for free because it makes money on the hardware — a structural advantage no pure AI company can match.
What This Means for Open-Source AI Agents
For the open-source agent ecosystem — frameworks like Hermes Agent, OpenClaw, CrewAI, and LangChain — Apple’s entry is a double-edged signal.
The bullish case: Apple is validating the entire category. When the world’s most valuable consumer technology company declares that AI agents are an OS-level feature, it accelerates adoption across the industry. Apple’s privacy-first, on-device architecture also aligns with the ethos of local-first open-source tools — many of which already run entirely on-device using quantized models. If Apple’s bet on local inference pays off, it lifts the entire local-agent ecosystem.
The bearish case: Apple is building a walled garden around agents. Siri Extensions — the framework for third-party AI models — is Apple’s distribution channel, and Apple controls the rules. If the App Store becomes the primary way users discover and install AI agents, open-source alternatives may struggle to reach consumers on iOS. The 30% commission that developers fear could apply to agent transactions, not just app purchases, creating a rent extraction model that open-source projects can’t afford.
The open-source community’s best defense is already in motion: build agents that run everywhere — browser, desktop, mobile web — and treat iOS as one platform among many, not the platform. Apple’s history suggests it will carve out a lucrative premium tier. The question is whether the open-source ecosystem can serve everyone else.
FAQ
Q: When does the new Siri actually ship? A: The first Gemini-powered features arrived in iOS 26.4 (spring 2026). The full Siri overhaul — standalone app, personal context, on-screen awareness — is expected to be announced at WWDC and ship with iOS 27 this fall.
Q: Will Siri use my data to train AI models? A: Apple’s architecture processes most requests on-device. Private Cloud Compute servers are ephemeral and verifiable — Apple publishes cryptographic attestations. Google does not receive user data through the Gemini partnership; the models run on Apple’s infrastructure.
Q: Can I use ChatGPT or Claude instead of the new Siri? A: Yes — Siri Extensions will let you select alternative AI backends. Each will use a distinct voice so you know which AI you’re talking to.
Q: Is this Tim Cook’s last WWDC? A: Yes. Cook is expected to hand the CEO role to John Ternus later this year. WWDC 2026 is widely seen as his valedictory keynote.
Q: Will AI agent apps actually be allowed in the App Store? A: Apple is designing a new framework for it, and an announcement at WWDC is possible. The timeline for developer access is unclear.
Further Reading
- What to expect from WWDC 2026 (TechCrunch)
- Apple picks Google’s Gemini to run AI-powered Siri (CNBC)
- WWDC 2026: What to Expect (MacRumors)
- Apple Working on Plan to Allow AI Agent Apps on the App Store (MacRumors)
- Apple’s $1B Gemini Deal (Tech-Insider)
- The Company Everyone Says Lost the AI Race Is Building the Layer Every AI Winner Has to Use (Nate’s Newsletter)
This preview is based on pre-conference reports, leaks, and announced partnerships. Final announcements may differ. We’ll publish a full analysis after the keynote.