Apple Will Let You Swap AI Models in iOS 27 — Siri With Claude or Gemini Plugged In

Apple just said, for the first time, "we won't lock down AI"

Here's the deal: on May 5, Bloomberg reported that iOS 27 / iPadOS 27 / macOS 27 — shipping this fall — will introduce an "Extensions" system letting users plug third-party AI models into every Apple Intelligence surface: Siri, Writing Tools, Image Playground. Beyond Apple's own Apple Intelligence Core LLM, users would be able to pick Anthropic Claude, Google Gemini, and others. Apple is in private talks with Anthropic and Google. This is the first time Apple has cracked the "closed ecosystem" doctrine in any meaningful way for AI — which means scenarios like "iPhone with Claude as the default voice assistant" become real within six months.

The cast — Apple, Anthropic, Google, OpenAI, users

Apple first. After WWDC 2024 launched Apple Intelligence and bolted ChatGPT on as a Siri backend, the integration was OpenAI-only. Apple's own LLM (estimated Apple Intelligence Core, ~30B parameters) trails GPT-5, Claude Opus, and Gemini 2.5 Pro by a wide margin. Through 2025 user complaints accumulated — "Apple Intelligence is the slowest and least accurate of the bunch." Extensions is Apple admitting the gap and routing around it through external models.

Anthropic Claude is the leading first-partner candidate. With the May 6 Anthropic-SpaceX Colossus 1 compute deal removing capacity constraints, Anthropic now has the headroom to serve Apple's 1.5B iPhones. Apple and Anthropic have been collaborating informally since 2024, and Anthropic's "safety-first" branding aligns naturally with Apple's "privacy-first" identity. Most likely first integration.

Google Gemini is also a strong candidate. Android's parent is also Apple's $20B+/year search-default partner, so the channels are deep. Gemini 2.5 Pro's multimodal stack (image, video, voice) brings outsized value to Vision Pro and Mac.

OpenAI ChatGPT is the incumbent integration partner; its position weakens under Extensions. The 2024 sole-integration ends and ChatGPT competes alongside others. That said, ChatGPT remains the most-used model — user-pick share could still be 50%+.

For Apple users (1.5B iPhones, 500M iPads, 200M Macs), the choice set widens. On privacy, all external model calls would route through Apple's Private Cloud Compute, minimizing data exposure. Model selection: Settings → Apple Intelligence → Default Model.

The hard numbers — Extensions structure, privacy, feature-to-model map

Structure and likely mapping.

Apple Intelligence feature	Default model	Extensions options
Siri conversation	Apple Core	Claude, Gemini, ChatGPT
Writing Tools (summarize / edit)	Apple Core	Claude (strong)
Image Playground	Apple Core	Gemini Imagen, DALL-E 3
Visual Intelligence	Apple Core	Gemini 2.5 Pro Vision
Code in Xcode	Swift Assist	Claude Code, GPT-Code
Notification summaries	Apple Core	Claude (strong)

Best-guess mapping by where each model has a real edge: Writing Tools and notification summarization to Claude, Image Playground to Gemini Imagen / DALL-E, Visual Intelligence to Gemini 2.5 Pro Vision.

The privacy structure is the most interesting piece. Apple is reportedly looking at routing external model calls through Private Cloud Compute, with user data nondeterministically transformed in the local keystore before any external API call. Anthropic and Google would not see raw user data. The trade-off: this likely degrades external-model quality, creating an "Apple privacy first vs external-model performance" tension to navigate.

Timing matters. iOS 27 GA is expected September 2026 with Extensions activated at launch. Formal announcement is highly likely at WWDC 2026 (June 9-13). Which model partners Apple puts on stage at WWDC sets the AI narrative for the next six months.

What each side gets — Apple, Anthropic & Google, users, developers

Apple gets three things. First, Apple Intelligence satisfaction recovery — the past 18 months of "Apple Intelligence is the worst of the AIs" gets routed around through external models. Second, hardware differentiation preserved — even if AI models come from outside, "iPhone Air 18 Pro camera + display + privacy" remains the differentiator. Third, an "Apple Tax" expansion — taking a fee per external model call could become a $3-5B/year incremental revenue stream.

Anthropic and Google get distribution: a direct path to 1.5B iPhone users. For Anthropic that's potentially Claude going from ~100M users today to 600M-1B reachable. The catch: Apple is likely to demand a 60%+ API discount plus ~30% revenue share — so think of it as +$5B ARR with margin compression, not pure upside.

Users get model choice. Medical questions to Claude, travel planning to Gemini, casual chat to Apple Core. The cost is more cognitive load — users have to learn which model is best for which task.

For iOS developers, Extensions is a single AI integration API. App Store apps can use the user's chosen default model directly — no separate Anthropic / Google / OpenAI SDK integrations needed. The iOS 27 SDK abstracts everything behind AppleIntelligence.shared.

Historical comps — App Store and the Default Browser change

Win #1: 2010 App Store launch. Apple flipping from "iPhone apps only by Apple" to "third-party developers welcome" remade the platform. Cumulative App Store revenue: $1T+. The first proof that Apple opening a door blows up the ecosystem. If Extensions follows that pattern, an "AI App Store" category gets born.

Win #2: 2022 default-browser change (DMA pressure). Apple let users set Chrome / Firefox as default. Safari share moved from ~85% to ~78% — modest erosion, big leap in user freedom. Same template for Extensions: Apple Core might fall from 50% to 30% share, but the ecosystem expands.

Bust comp: 2014 Apple Music + Beats. Apple bought Beats for $3B and launched Apple Music — yet Spotify won on share. Hardware + software integration alone wasn't enough. Apple Intelligence post-Extensions could end up in the Spotify pattern: Apple's own model is weak, and external models do most of the work.

Balanced comp: 2018 Apple Silicon transition (closed: only Apple chips for Mac). Mac revenue $25B (2018) → $38B (2025), +50%. The opposite vector to Extensions (closed → open) — but both moves work because Apple knows where its leverage actually sits.

One more: 2023 ChatGPT Plugins. OpenAI opened ChatGPT to third-party app connections. Mixed: Wolfram and Expedia plugins were popular, but plugin usage stayed under 10% of total ChatGPT time. Lesson: Extensions on its own doesn't shift behavior. Apple Intelligence Extensions could see the same: most users stick with default unless prompted.

Counter-plays — Google Pixel, Samsung Galaxy AI, Microsoft Copilot+ PC

Google Pixel pushes "AI-first phone" with Gemini Nano on-device + Gemini Pro cloud. If iOS 27 Extensions adopts Gemini, Pixel's Gemini differentiation weakens. Mixed for Google: capturing Apple distribution is a win, weakening Pixel differentiation is a cost.

Samsung Galaxy AI uses a multi-model approach: Gemini + Bixby + Samsung's own. Samsung announced "Galaxy AI Plus" (user-pickable Claude / OpenAI / Gemini) on May 8 — within hours of the Apple report. The Apple-Samsung "AI model freedom" race has officially started.

Microsoft Copilot+ PC pairs Windows 11 with Snapdragon X Elite + Copilot. macOS 27 Extensions on Mac with Claude / Gemini integration puts Apple in direct competition there. May 7 reports indicate Microsoft is exploring "Copilot+ PC Extensions" (third-party model integration) in response.

xAI Grok rides Tesla integration. Whether iOS Extensions includes Grok is open — the Tim Cook-Elon relationship is delicate, but on user-choice grounds Grok inclusion is plausible.

China — Doubao, Kimi, DeepSeek, Qwen. iOS Extensions prioritizes Western markets first; mainland China iOS would likely get a separate set of Chinese model integrations. Chinese users can't use ChatGPT / Claude / Gemini, so a parallel path is required.

So what changes — users, developers, LLM company strategy

For Apple users, "model literacy" becomes a thing. iPhone users will need to figure out which model fits which task. Auto-suggestions ("Claude is faster for this") are likely; the exact UX gets revealed at WWDC and iOS 27 launch.

For iOS developers, "AppleIntelligence API" becomes the standard. Today, integrating each of ChatGPT / Claude / Gemini SDKs separately is overhead; from iOS 27, a single AppleIntelligence.complete() call routes to the user-chosen default. Estimated AI-integration cost reduction: 50-70%.

For LLM companies, "Apple Tax" creates a new cost line. Apple is likely to demand 30% per-call fees or ~30% ARR share on Extensions integration. ChatGPT / Claude / Gemini each face the choice: accept Apple Tax to grab distribution, or refuse and double down on direct channels. Likely outcome: Anthropic and Google accept; OpenAI negotiates down to 15-20%.

For enterprise IT, BYOD changes meaningfully under iOS 27 + Extensions. The key question: when employees use Claude or Gemini on their iPhone, how is corporate data protected? Apple plans MDM-level "company-owned model only" policies admin-configurable.

The biggest macro signal: Apple ending exclusive OpenAI integration resets distribution averages across the LLM market. The "1.5B iPhones for ChatGPT only" advantage disappears, and Anthropic / Google / xAI all gain comparable reach. That's the largest variable for the next 12-24 months of LLM market share. The current ChatGPT 800M / Claude 100M / Gemini 400M gap could close meaningfully.

References

• iOS 27 Features: Apple Plans to Let Users Swap Models Across Apple Intelligence — Bloomberg
• Apple Intelligence Overview — Apple
• Private Cloud Compute Security — Apple Security Research
• WWDC 2026 (preview) — Apple Developer
• App Store Ecosystem Statistics — Apple Newsroom