Apple Intelligence vs ChatGPT vs Gemini: a 2026 comparison of Apple’s AI models

The three names are not on the same layer

Split product, model, and API first. If they share one cell, the “full comparison” will compare the wrong objects.

Apple Intelligence is the system layer. It covers Writing Tools, Image Playground, notification summaries, and the 2026 rebuild of Siri AI. Users see Siri and system features, not a model trademark. Next-gen Apple Foundation Models are custom-built with Google using Gemini-family technology, but the experience stays on Apple Intelligence; Craig Federighi’s line is still that the amount of Google Assistant they use is none. Light work stays on-device; heavier requests go to Private Cloud Compute — including the June tier that landed on Google Cloud. How to read the strategy: why Apple is leaning harder on Gemini.

ChatGPT is OpenAI’s consumer product and a developer API. Since 2025 it has sat inside Siri as an Apple Intelligence extension: hard questions can go to OpenAI after you confirm. Apple’s support page still shows how to turn ChatGPT on under Apple Intelligence & Siri in 2026. It is a side door, not the AFM training base. Reports described a multi-vendor Extensions framework in the iOS 27 betas; as of September 2026 Apple has not announced it at WWDC, and the only public extension is still ChatGPT. Gemini occupies three cells at once: Google’s consumer assistant, a developer API with Structured Output, and Apple’s infrastructure deal. One word, three JSON dialects, three privacy fences.

What actually shipped for each in 2026

Start with the official timeline. The 12 January joint statement: a multi-year collaboration; next-gen Apple Foundation Models based on Gemini models and cloud technology; future Apple Intelligence, including a more personalized Siri; experience still on-device and on PCC. On 8 June two facts stacked: Apple Security Research published Expanding Private Cloud Compute — PCC reaches third-party halls for the first time, and the heaviest agent tool-use runs on NVIDIA GPUs in Google Cloud — and the same week’s WWDC26 introduced Siri AI: on-screen awareness, personal context across apps, systemwide App Actions, plus a dedicated Siri app that syncs history over iCloud. On the developer side the Foundation Models framework grew from on-device-only into one Swift API that can also hit server models, and opened a LanguageModel protocol; Apple named Claude and Gemini.

On the ChatGPT side, 2026 did not “uninstall OpenAI.” Support docs still describe an optional extension: Siri asks before it helps, and photos or files need confirmation; “Ask ChatGPT” skips the prompt. What changed is weight. System Q&A and App actions are increasingly caught by Siri AI + AFM; ChatGPT looks more like a spare exit for world knowledge and long reasoning. OpenAI’s own flagship, in our 1M-token comparison, is GPT-5.6 Sol / Terra: about 1.05M window, 128K output. That is the model card when you call the OpenAI API — not a window Siri’s extension promises.

On the Gemini side, split “for Apple” from “for you.” For Apple it is AFM teacher signal / foundation and the PCC expansion; Settings has no Gemini toggle. For you it is the Gemini API, plus the Firebase preview that plugs into Foundation Models: swap SystemLanguageModel for GeminiLanguageModel and @Generable plus Tool definitions can stay — but the request leaves PCC, and Apple’s five privacy requirements no longer apply. Reported figures (about a billion dollars a year, a custom ~1.2T-parameter Gemini) are unconfirmed in the joint statement. Treat them as scale color, not an architecture input.

Seven axes side by side

The table uses public facts you can still align in September 2026. It compares default paths, not “who wrote a better essay in one test.” Essay scores drift with the prompt and the date. Inference location, privacy fence, and developer entry do not.

When you read the table, do not flatten “Gemini appears in three columns” into one request. AFM using Gemini tech does not mean Siri’s HTTP becomes generateContent. A ChatGPT extension leaving the device does not rewrite your App Intents parameters into OpenAI tool_calls. All three columns can be true at once because they are not the same layer.

Which jobs each one wins

For end users. On an Apple device, “turn the address in this mail into a reminder” or “put the itinerary on screen onto the calendar” is Apple Intelligence / Siri AI: it sees the screen and personal context, and it can hit system App Actions. Long specs, a large code edit, or the same thread across platforms is usually more stable in ChatGPT — window and max output follow the model card, and you are not boxed by Siri’s English-first rollout or the EU iPhone delay. Video or PDF in a Google-account life is where Gemini’s multimodal window is wider. On EU iPhones Apple said Siri AI would not land first on iOS / iPadOS / watchOS, while Mac and Vision Pro can; the phone comparison there is still old Siri plus the ChatGPT extension.

For developers the question is narrower: who fills your arguments. When the system agent hits your App in 2026, it still arrives through App Intents — typed parameters, not a blob of Gemini prose. In-app on-device or PCC models still use Foundation Models. If your own backend wants a frontier cloud model, pick between the OpenAI API and the Gemini API — another bill, another set of field names. The LanguageModel protocol lets you swap a Claude or Gemini instance in Swift. Swapping the instance does not retire App Intents and does not inherit PCC. How the three call chains split: how an Apple AI Agent reaches Apps and APIs.

A “full comparison” has no single winner. The stack that wins device control loses the open ecosystem. The 1M window does not win the system calendar. The privacy story does not win “keep chatting on any web page.” Cut by job, not by trademark.

Do not mix the three JSON chains

The contract you control is still the shape, not the model trademark. When system Siri AI hits your App, arguments come from App Intents. In-app models build Arguments with @Generable; you decide in call(arguments:) whether to JSON-encode an HTTP body. You only see json_schema or responseSchema when your server calls OpenAI or Gemini directly. Cloud models in 2026 are bolder about multi-step tools, so extra keys and type drift show up more than in 2025 — we said that in the Gemini strategy piece, and a three-way comparison does not change it.

Store golden samples at the intent layer. The JSON below only says who came in, which intent, and which arguments. ingress.surface is fine in logs. Do not put it in business validation. Siri AI, a ChatGPT result you forwarded yourself, and the Gemini API must accept the same arguments Schema, or one fallback or extension swap will break order lookup.

{
  "intent": "findOrders",
  "arguments": {
    "email": "ada@example.com",
    "status": "paid",
    "limit": 5
  },
  "ingress": {
    "surface": "siri-ai",
    "adapter": "app-intents"
  }
}

Once you have live traffic, drop pass and fail samples into JSONVue: format to see fields, JSON Schema to pin types and enums, Diff to see which keys the on-device fallback drops versus Cloud Pro / an external API. Structured Output pitfalls for ChatGPT and Gemini are in AI Structured Output and Gemini API JSON output; window budgets are in the 1M-token comparison. Model trademarks will move. The JSON shape should not.

FAQ

Summary and next steps

The 2026 comparison compresses to one sentence: Apple Intelligence wins the device and the privacy fence, ChatGPT wins open-domain reasoning and the API ecosystem, and Gemini powers both Google’s product line and Apple’s model base. All three names can sit on one iPhone. They are not one request.

The next step is concrete: write App Intents parameters and HTTP tool payloads as one Schema, validate pass/fail samples in JSONVue, and assume Cloud Pro or an external API will fill two extra optional keys. Align the shape first. Argue about trademarks later.