Review: Apple Vision Pro 2 — The First True Spatial Intelligence Device

There is a moment, roughly forty-eight hours into living with the Apple Vision Pro 2, when the device stops feeling like a product and starts feeling like a condition. You are standing in your kitchen, a translucent recipe card floating at eye level, the morning light bleeding through your actual window mixing with the synthetic glow of a notification from someone in Tokyo. The boundary between the room you are standing in and the room Apple has built inside your skull has become, for a few seconds, genuinely ambiguous.

That ambiguity is the product. Everything else — the M5 Ultra chip, the micro-OLED panels running at 8K per eye, the 14 cameras mapping your environment in real time, the eye-tracking so precise it can distinguish between reading and scanning — is infrastructure in service of that ambiguity. Apple has not shipped a headset. They have shipped a new sensory layer for human cognition, and after thirty days inside visionOS 3, the verdict is unambiguous: the Vision Pro 2 is the most consequential computing device since the original iPhone. And the most unsettling.

The Hardware Is No Longer the Story

The original Vision Pro, released in 2024, was a proof of concept wearing a price tag. Extraordinary engineering, software that felt like a beta, and a form factor that announced itself in every room you entered. The Vision Pro 2 is different in kind, not just degree. The chassis has shed 40% of its mass. The battery is now integrated into a modular spine that distributes weight across the crown of the skull rather than cantilevering it off your face. You can wear it for four hours without the low-grade headache that made the first generation feel like a punishment for curiosity.

The display is, without qualification, the finest screen ever placed in front of human eyes. Apple's custom micro-OLED panels render at a pixel density that exceeds the resolving power of the human fovea at normal viewing distances. In practical terms: you cannot see the pixels. You can read 6-point type floating three feet in front of you. Text in the Vision Pro 2 is sharper than text on paper.

The M5 Ultra chip processes the sensor array — cameras, LiDAR, depth sensors, microphones — at a latency Apple claims is under 12 milliseconds. The claim holds up. When you reach for a virtual object, your hand arrives before your brain has finished predicting where it should be. The system is faster than your nervous system's own feedback loop, which produces a sensation that is difficult to describe and impossible to unfeel once you have experienced it.

visionOS 3: Where the Intelligence Lives

The hardware is remarkable. The software is where Apple has made its real bet.

visionOS 3 introduces what Apple calls Spatial Intelligence — a persistent, on-device AI layer that builds a semantic model of your physical environment and your behavioral patterns simultaneously. It knows that the corner of your desk near the window is where you do focused work. It knows that you read long-form content in the evening and scan communications in the morning. It knows, with a precision that should probably alarm you more than it does, the difference between the face you make when you are concentrating and the face you make when you are bored.

This is not cloud inference. The entire model runs on-device, on the Neural Engine of the M5 Ultra, which Apple has expanded to 64 cores specifically to support continuous environmental reasoning. The privacy architecture is, by current standards, genuinely serious: no sensor data leaves the device, the semantic model is encrypted and stored locally, and a hardware kill switch physically disconnects the camera array when the device is in Mindful Mode. Apple has read the room on surveillance anxiety, and they have responded with engineering rather than policy.

The practical result of Spatial Intelligence is a device that feels, after a few weeks of use, less like a tool you operate and more like an environment you inhabit. Windows arrange themselves according to your attention patterns. Notifications surface at the periphery of your vision during low-focus states and suppress themselves entirely when the system detects deep work. Your most-used applications migrate toward the positions in space where you most naturally look for them. The device is learning the geometry of your attention, and it is doing so without asking permission or announcing its conclusions.

The Photonic Question

Here is where the Vision Pro 2 becomes interesting to anyone tracking the longer arc of computing infrastructure. The device runs entirely on conventional silicon — the M5 Ultra is a TSMC 3nm part, fast and efficient and fundamentally unchanged in architecture from the chips that have powered Apple devices for the past decade. But the display pipeline, the sensor fusion layer, and the neural rendering engine all operate at the boundary of what silicon can sustain at this power envelope.

The Vision Pro 2 is, in a meaningful sense, the last great silicon spatial computer. The next generation of this device — the one that will be lighter than glasses, that will run for a full day, that will render environments indistinguishable from physical reality — will require photonic interconnects to move data between the sensor array and the processing layer at the speeds the application demands. The latency wall that silicon hits at this form factor is not an engineering problem that more transistors will solve. It is a physics problem that light will solve. For a deeper look at where that transition is heading, see our Photonic Computing pillar.

The Vision Pro 2 makes this legible in a way that no spec sheet can. You feel the device working. In complex mixed-reality scenes — a full spatial workspace with a dozen active windows, live video feeds, and a 3D model being manipulated in real time — there is a thermal signature, a fan curve, a subtle increase in the weight of the device as the battery depletes faster than the rated curve. The M5 Ultra is running at its ceiling. The next ceiling is photonic.

The Identity Problem

Philip K. Dick asked whether androids dream of electric sheep. The Vision Pro 2 asks a quieter, more personal question: whose memories are these?

After thirty days, the device has accumulated a spatial autobiography. It knows the layout of every room I work in. It has mapped the faces of the people I spend time with — locally, on-device, never uploaded, but mapped nonetheless. It has built a model of my attention that is, in certain respects, more accurate than my own self-report. When I asked it to show me my focus patterns for the past week, the visualization it produced was both illuminating and faintly accusatory.

This is the territory that the AI identity discourse has been circling without quite landing on. The threat to human identity from artificial intelligence is not replacement — it is not that a machine will do your job better than you. The threat is accumulation. The slow, consensual, entirely voluntary construction of a parallel self inside a device that knows you better than you know yourself, and that will, eventually, outlast you. We explored this question directly in our piece on Ghost in the Shell and AI identity — the 1995 film remains the most accurate map of this territory.

The Vision Pro 2 is the first consumer device that makes this dynamic visceral rather than theoretical. You are not using it. You are being modeled by it. The distinction matters, and Apple, to their credit, has built the privacy architecture to make that modeling at least nominally yours. But the architecture is only as durable as the company that maintains it, and companies change.

For the academic framing of where on-device AI identity modeling is heading, the Anthropic alignment research team and the AI Now Institute are publishing the most rigorous work on the governance questions this device raises. Apple's own Machine Learning research blog documents the on-device inference architecture in unusual technical detail for a consumer company.

The Verdict

The Apple Vision Pro 2 is the finest piece of consumer technology ever manufactured. The display is without peer. The spatial intelligence layer is genuinely novel. The privacy architecture is, for now, trustworthy. The form factor, while still not invisible, is no longer a social liability.

It is also the most philosophically loaded object you can put on your face. It is a device that will change how you perceive physical space, how you relate to digital information, and — if you let it — how you understand the structure of your own attention. These are not small changes. They are the kind of changes that, once made, cannot be unmade.

The question is not whether the Vision Pro 2 is worth $3,499. For anyone who works in knowledge, creativity, or spatial reasoning, it is. The question is whether you are ready for a device that will know you better than you know yourself, and whether you trust the company that built it to steward that knowledge with the care it deserves.

The neon-noir future that cyberpunk fiction spent forty years warning us about did not arrive as a dystopia. It arrived as a product. It arrived beautifully designed, impeccably engineered, and priced for the professional class. It arrived, as these things always do, looking exactly like progress.

Apple Vision Pro 2 — $3,499. Available now. Reviewed on visionOS 3.0.1 over 30 days of daily use.