Everything AI Apple Announced at Its September 9 Event

Apple just used John Ternus’s first major keynote as CEO to answer a question hanging over the company for years: what does Apple’s AI strategy actually look like when it ships?

The answer is less “one giant chatbot” and more AI threaded through Siri, the operating system, the camera, Shortcuts, Safari, the silicon inside the iPhone, the manufacturing process behind the iPhone Duo, and the sensors on your wrist.

There is one important caveat. A lot of the software Apple showed today was first previewed at WWDC in June. The September event was more about turning those promises into shipping products, giving Siri AI a release window, and pairing it with hardware designed to run more AI locally.

Apple also published unusually detailed information about the models underneath all of this. That part is easy to miss if you only watched the hardware keynote.

That said, if you missed it, you can watch the full thing here.

Here’s the AI news that actually matters.

1. Siri AI is finally moving from demo to product

Siri AI is not a finished stable release yet. Apple says the English beta rolls out with iOS 27 on Monday, September 14, with French, Japanese, Korean, Portuguese, and Spanish coming in October. Apple has not announced a separate date for a non-beta release. Some compute-heavy Apple Intelligence features carry daily usage limits.

This is a much bigger Siri rebuild than a new voice or UI. Siri AI can use personal context from messages, emails, photos, and other apps, understand what is on your screen, pull current information from the web, and take actions across apps.

Apple also built a dedicated Siri app that lets you revisit old conversations or start new ones. Conversation history can privately sync through iCloud across Apple devices.

The bigger idea is that Siri is becoming less like the thing you ask to set a timer and more like an AI layer over the entire operating system.

2. Apple now has five third-generation foundation models

The most important technical detail did not come from a giant benchmark slide. It came from Apple’s Machine Learning Research team.

Apple’s third-generation model family, introduced at WWDC in June, contains five models custom-built in collaboration with Google:

• AFM 3 Core: Apple’s next-generation dense on-device model, with roughly 3 billion parameters.
• AFM 3 Core Advanced: Apple’s most powerful on-device model. It is natively multimodal and has 20 billion total parameters, but a sparse architecture activates only 1 to 4 billion at a time depending on the request.
• AFM 3 Cloud: Apple’s server-side workhorse for more demanding multimodal reasoning, optimized for speed, efficiency, and performance.
• ADM 3 Cloud (Image): Apple’s image-generation and editing model for Image Playground, Genmoji, photo editing, Spatial Reframing, and related visual features.
• AFM 3 Cloud Pro: Apple’s most capable server model, designed for complex reasoning and agentic tool use.

Four of those models are optimized for Apple silicon. AFM 3 Cloud Pro is the exception: Apple says it worked with Google and NVIDIA to run the model on NVIDIA GPUs in Google Cloud while extending Private Cloud Compute’s privacy protections to that infrastructure.

AFM 3 Core Advanced is particularly clever. Rather than trying to keep all 20 billion parameters in active memory, Apple stores the full model in flash and loads only the experts needed for a request into DRAM. That lets the phone use a much larger total model without needing 20 billion parameters active for every task.

In plain English: Apple built one on-device model that can effectively scale its active size with the difficulty of the job.

3. Apple’s benchmark story is strong, but different from OpenAI or Google’s

Apple published detailed evaluations for these models alongside its June model rollout, although it is important to read them correctly.

The current 2026 research page does not give us a clean head-to-head chart against GPT, Gemini, or Claude on the usual public benchmark suite. Instead, Apple mostly reports human preference evaluations against its own previous-generation systems.

• AFM 3 Core: preferred on 45.6% of general-text prompts versus 23.3% for Apple’s 2025 baseline. On image understanding, it was preferred over the prior generation more than 61% of the time when graders chose a winner.
• AFM 3 Cloud: preferred on 64.7% of text prompts versus 8.7% for the 2025 server model. Apple also reports roughly 36% better overall response satisfaction and 21% better instruction following.
• AFM 3 Cloud Pro: improves another roughly 10% over AFM 3 Cloud on text satisfaction, 14% on image understanding, and 14% in Apple’s Math category.

The feature-level numbers are easier to picture. At a 1-billion-active-parameter configuration, AFM 3 Core Advanced scored 4.24 out of 5 for conversational voice quality versus 3.82 for Apple’s prior text-to-speech system. In dictation, graders preferred the new system 44.7% of the time versus 17.6% for the old one.

Apple’s separate audio research says the expressive-voice system runs about 16 times faster than real time on Apple’s Matrix Coprocessor while using roughly 21MB of runtime memory.

So Apple can credibly say its models took a large generational jump. What it cannot credibly claim from these published numbers is that AFM 3 Cloud Pro beats the latest frontier models from OpenAI, Google, or Anthropic. Apple has not published that comparison here.

4. Siri is moving into the camera, writing, calls, and dictation

Apple is also expanding where you can use Siri AI.

With Siri mode in Camera and Visual Intelligence, users can point the camera at something, ask questions about what they see, and trigger actions. Siri can also help write or edit text across the system rather than requiring users to open a separate chatbot.

AFM 3 Core Advanced powers more expressive Siri voices and higher-accuracy dictation on supported hardware. Users can adjust Siri’s pace and expressiveness, while dictation can better handle punctuation, capitalization, formatting, and natural speech.

Two smaller features are worth calling out because they show Apple’s system-level advantage:

• Call Context can proactively surface information from other apps during a business call, such as pulling an airline confirmation code from Mail while you are speaking to the airline.
• Suggestions in Messages and Mail can use conversation context to offer actions such as creating an event, adding a reminder, or finding a photo.

These sound less dramatic than a frontier-model benchmark. They are also exactly the sort of AI features people could end up using dozens of times a week without thinking of them as “using AI.”

5. Apple is turning everyday apps into little agents

Some of Apple’s most interesting AI features do not look like AI apps at all.

Safari’s Notify Me can monitor a webpage for a change, such as a product restock or price drop, and alert you when it happens. Safari can also automatically group tabs by topic, while AI-assisted extensions can personalize webpage content or automate common browser actions.

Describe a Shortcut lets you explain an automation in normal language and have Shortcuts assemble the steps across apps.

Apple is also exposing its on-device foundation model to developers through the Foundation Models framework. Apple says apps can use that model offline with no per-request model fee, which is a very different economics model from calling a paid cloud API every time an app needs AI.

The direction is more consequential than “Siri can answer harder trivia.” Apple is taking actions people already perform manually across apps and letting AI connect the steps.

6. The A20 Pro may be Apple’s most important AI hardware announcement

The iPhone 18 Pro and Pro Max run on Apple’s A20 Pro, built around a new architecture designed for heavier local AI.

When can you actually buy them? Preorders open at 5 a.m. PT on Saturday, September 12, and iPhone 18 Pro and Pro Max become available Friday, September 18 in more than 65 countries and regions. Another 20 markets follow September 25.

Apple’s own product page calls it a “total AI powerhouse.” The chip combines new Neural Accelerators with a Dual 16-core Neural Engine that Apple says delivers 2x the compute for on-device intelligence, plus 50% higher memory bandwidth.

That memory bandwidth matters because large models need to move enormous amounts of model data quickly. More bandwidth can be just as important as raw compute when you are trying to make a model respond instantly on a phone.

Apple also redesigned the chip packaging and thermal system. The new vapor chamber has 3x the surface area, and Apple says the combined design delivers up to 40% better sustained performance than iPhone 17 Pro.

Apple explicitly connects that thermal gain to AI: the phone can run larger, more complex AI models on-device for longer before heat becomes the limiting factor.

This is Apple’s core AI advantage if it works: do more locally, send less personal data to a remote server, and make AI feel like part of the phone rather than a website you visit.

7. The iPhone Duo uses AI before you even turn it on

Apple’s first foldable, the iPhone Duo, adds a completely different AI story: Apple used AI to help manufacture the hardware itself.

Duo comes later: preorders open at 5 a.m. PT on Friday, October 16, with availability starting Friday, October 23 in more than 70 countries and regions. Another 28 markets follow October 30.

One of the easiest details to miss in the keynote was Apple saying it used AI in the manufacturing process to tune the Duo’s hinge components so the paired parts work together precisely. That is AI being used as a manufacturing-control tool, not as a chatbot feature.

The Duo itself opens from a phone-sized outer display into the largest display Apple has put on an iPhone, with side-by-side multitasking and Apple Pencil support also highlighted in launch coverage.

It also runs the A20 Pro, so it gets the same emphasis on advanced local AI workloads as the Pro line. Apple did not announce a special Duo-only foundation model or assistant.

The Duo’s AI pitch is really the combination of AI-assisted manufacturing, top-tier on-device compute, and a larger canvas where system-level AI can work alongside multiple apps.

8. Apple is pairing AI creation with proof that a photo is real

The iPhone 18 Pro adds provenance tools at a moment when AI-generated imagery is getting harder to distinguish from camera photos.

Apple’s new Apple Reference Image system securely signs photo data at the pixel level at capture, creating a way to verify what the camera sensor actually saw. Apple also highlighted support around identifying AI-generated content during the keynote.

That creates an interesting two-sided strategy: Apple wants to make AI image creation and editing more powerful while also giving users a way to establish provenance for real camera captures.

Meanwhile, Apple Intelligence can reframe a photo after capture with Spatial Reframing, extend the edges of an image, remove larger objects with Clean Up, and generate or edit images in Image Playground. Smart focus tracking also uses on-device intelligence to keep a subject in focus even if the subject leaves and reenters the frame.

Underneath the generative side is ADM 3 Cloud (Image), Apple’s dedicated cloud image model. Apple says it handles image creation, editing, and Genmoji, while specialized adapters power features such as Spatial Reframing.

9. Apple’s AI architecture is increasingly hybrid

Apple is not trying to run every AI request on one giant model in one place.

Its stack now combines small on-device models for fast tasks, a 20-billion-parameter sparse on-device model for heavier local work, multiple Private Cloud Compute models, and a Cloud Pro model running on NVIDIA GPUs in Google Cloud.

On top of those models sits Apple’s system orchestrator, which can tap into local components such as the Spotlight index and App Toolbox to retrieve context and take actions.

That architecture explains why Apple is pouring so much attention into silicon. The better the local models and hardware get, the more requests Apple can handle without sending sensitive context elsewhere.

It also makes Apple’s Google partnership more nuanced than “Siri runs on Gemini.” Apple says these are Apple Foundation Models custom-built in collaboration with Google, with Apple controlling the product architecture and privacy layer.

10. Apple Watch is becoming an AI memory layer

The Apple Watch Series 12 may contain Apple’s most unusual AI product of the day.

Audio Intelligence combines the Watch microphone, S11 chip, Apple Intelligence models on iPhone, and Private Cloud Compute:

• Sound Recognition listens for important sounds such as sirens, alarms, doorbells, and a baby crying, then uses on-device intelligence to alert the wearer.
• Live Rewind lets you double-press the Digital Crown and recover the previous 15 seconds of a conversation as text. You can ask Siri about it or save the snippet.
• Siri Recap can generate a title and high-level key points after a conversation when the user has opted in.
• Shazam can automatically identify music playing around the wearer and surface it in the Smart Stack.

The privacy architecture is as interesting as the feature. Apple says Audio Intelligence does not create or store audio recordings. The S11 chip includes a Secure Exclave, a hardware-isolated compartment that processes raw audio and immediately deletes it.

Audio Intelligence also does not identify or attribute speakers. Saved Live Rewind snippets and Siri Recap summaries are end-to-end encrypted through iCloud.

And with watchOS 27, Siri AI itself comes to Apple Watch, bringing personal context and broad world knowledge to the wrist.

That makes the Watch look less like a tiny phone and more like a memory layer you happen to wear.

11. Apple is using AI to turn health sensors into decisions

Apple also gave a much clearer answer to what “AI for health” means on its devices.

The new Series 12 Health Sensing System measures heart rate every five seconds and samples heart-rate variability up to 24 times more frequently. Apple then uses health signals, activity, training load, vitals, and sleep to generate a new Readiness score from 0 to 10 with recommendations such as Recover, Pace Yourself, Ready, or Go For It.

The redesigned Health app uses Apple Intelligence to surface personalized insights. A new Longevity tab includes Health Age, which compares health indicators with a user’s chronological age.

Apple is also using vision-based AI models on iPhone plus Watch heart-rate data for movement evaluations covering flexibility, strength, balance, and more.

There is another easy-to-miss machine-learning upgrade: Apple says the Series 12 pedometer model was completely redesigned with new ML algorithms, and claims the result is the most accurate step tracking of any smartwatch.

This builds on years of health ML work. Apple’s hypertension model was developed with data from studies totaling more than 100,000 participants, its workout motion model has been trained on 50 million hours of activity data, and its sleep-score work used more than 5 million nights of study data.

The transferable idea is simple: sensors produce messy streams of numbers. AI turns those streams into a signal a normal person can act on.

12. AirPods extend Apple’s AI interface, but Live Translation is not brand new

AirPods showed up in Apple’s intelligence story mainly as another hands-free interface for Siri and translation.

Live Translation is powered by Apple Intelligence, but it is important not to treat it as a brand-new September 2026 invention. Apple had already brought Live Translation to supported AirPods before this event.

So the AI story here is not a special AirPods foundation model. It is Apple making the same system-level intelligence easier to access through more hardware.

What Apple actually announced today vs. what was already known

The cleanest way to read the event is:

• June was the software and model blueprint: Siri AI, personal context, Visual Intelligence, natural-language Shortcuts, the third-generation Apple Foundation Models, and Apple’s new AI architecture.
• September was the productization: new iPhone hardware built around heavier on-device AI, AI-assisted iPhone Duo manufacturing, Siri AI placed front and center in the iPhone pitch, and a Watch designed around Audio Intelligence and AI-generated health guidance.
• The model details matter more than Apple emphasized onstage: Apple now has five named third-generation models, including a 20B sparse on-device model and Cloud Pro running on NVIDIA GPUs through Google Cloud.
• Some headline features were extensions, not debuts: AirPods Live Translation already existed on supported models before this event.

That distinction matters. Apple did not suddenly unveil an entirely new AI strategy today. It showed how the strategy it previewed in June is supposed to become an everyday product, and how deeply AI is now moving into Apple’s hardware, manufacturing, health, camera, and operating-system layers.

The next test is not a benchmark chart. It is whether Siri AI can reliably do useful things across a person’s actual apps, messages, photos, routines, conversations, and health data without becoming another assistant people stop using after a week.

Apple’s own evaluations show a substantial generational jump. Now the September 14 beta has to prove that those gains survive contact with billions of real-world devices.