Google opened Pixel 11 preorders on August 12 and put the phones on shelves August 20, with the Pixel 11 starting at $899, the Pixel 11 Pro at $1,099, and the Pixel 11 Pro XL at $1,299. The headline is the new Tensor G6 — but the most interesting thing about it is the spec Google did not ship.
The Tensor G6 Is Not the 2nm Chip Everyone Predicted
In mid-July, reporting out of Taiwan claimed TSMC would build the Tensor G6 on its 2nm process, and the claim spread far enough that plenty of launch-day coverage repeated it as fact. It was wrong. Google's VP of hardware confirmed at the Taiwan launch event that the G6 is built on TSMC's N3P — a refined 3nm node, the same node generation as the Tensor G5 in the Pixel 10.
That is worth sitting with, because process nodes have become a marketing proxy for "is this chip good," and they are a bad one. Staying on N3P means Google spent its engineering budget on architecture and accelerator design rather than on being first to a new node — and on a phone SoC, where thermals and battery dominate the experience, that is a defensible trade.
What Google Actually Improved
The gains Google claims over the G5 are concentrated in the AI accelerator rather than raw CPU throughput:
- ›50% more TPU compute, paired with the latest Gemini Nano model
- ›On-device AI tasks running up to 3.5x faster while using up to 3.5x less energy
- ›An upgraded CPU delivering 25% faster web browsing
- ›15% quicker app launches
- ›Pro models reaching up to 3,600 nits peak brightness with more than double the scratch resistance
- ›Pixel 11 Pro XL charging to roughly 15 hours of battery life in 15 minutes on a wired charger
Treat those as vendor numbers until independent benchmarks land — they are Google's figures, measured Google's way. But the shape of the claim is credible: a 50% larger TPU on the same node is exactly the kind of thing you get by spending die area on the accelerator instead of chasing a shrink.
The efficiency figure matters more than the speed figure. A model that runs 3.5x faster is a demo; a model that runs on 3.5x less energy is a feature you can leave switched on all day.
Gemini Intelligence and the Agentic Turn
The software story is the bigger shift. Gemini Intelligence on the Pixel 11 is being positioned as an agent rather than an assistant: Google demonstrated it booking rides and ordering groceries by operating across third-party apps on the user's behalf, not just answering questions about them. Circle to Search also moved into the camera viewfinder, so you can identify an object or translate text in front of you without taking a picture first.
This is the direction every phone maker is walking, and it raises the same unresolved questions everywhere: what happens when the agent buys the wrong thing, and how much of your app session state are you comfortable handing to a model? Worth watching how Google handles confirmation steps in practice.
The Camera Is Still the Reason to Buy
Pixel has always been a computational photography story, and the Pixel 11 leans harder into it:
- ›A larger 48-megapixel main sensor on the Pixel 11 with 56% more light sensitivity, plus 30x Super Zoom
- ›An upgraded 48-megapixel telephoto on the Pro models with 30% more light sensitivity, Portrait Mode at 5x, and Pro Zoom to 120x
- ›Magic Capture, which analyses roughly 400 frames across a shooting session and combines on-device processing with Gemini models to pick a well-timed 12-megapixel frame
- ›Night Sight capturing up to 4.5x faster than previous generations
Magic Capture is the clearest example of what the bigger TPU is for. Sifting 400 frames and reasoning about which one is the good one is not a task you send to a server — the latency and the upload cost would both be absurd. It only works because the silicon underneath got wider.
What This Means If You Are Building
For anyone shipping mobile or IoT work, the practical takeaway is that on-device inference has quietly become a real deployment target rather than a research demo. A phone that can run a capable small model continuously, at low energy cost, changes what you can assume about connectivity in your architecture. Offline-first stops being a fallback mode and starts being a design default — and that is a more durable shift than any single spec on the box.
Written by Manas Garge
Founder & Data Engineer