Tensor G6: on-device AI 3.5x faster and off your bill

Google announced the Pixel 11 line at Made by Google today, and the number worth your attention is not the camera bar. Tensor G6 packs 50% more TPU compute and runs on-device AI tasks up to 3.5 times faster while using up to 3.5 times less energy. That is on-device AI crossing from demo to production — and every task it absorbs is a task that stops billing you per token.

What actually happened

Google's own announcement puts it directly: "Packing 50% more TPU compute and paired with the latest Gemini Nano model, Google Tensor G6 processes on-device AI tasks up to 3.5 times faster while using up to 3.5 times less energy." The chip also claims 25% faster web browsing and 15% quicker app launches from an upgraded CPU.

The lineup: Pixel 11 at $899, Pixel 11 Pro at $1,099, Pixel 11 Pro XL at $1,299, pre-orders open today. 9to5Google's hands-on coverage adds a Pixel Watch 5 and a UWB Pixel Tag tracker.

Two things we are deliberately not asserting. Google's post says "the latest Gemini Nano model" without a version number — the "Nano 4" label is coming from press coverage, not the vendor. And the 2nm process claim circulating before the event does not appear in Google's announcement, so treat it as unconfirmed.

Why on-device AI matters for your business

The 3.5x figures are relative to last generation, and vendor benchmarks are vendor benchmarks. But the direction is the point, and it has been consistent for three chip generations now: the class of AI work that runs locally keeps getting bigger, and it runs at zero marginal cost.

Look at what your field team actually asks a model to do. Transcribe a site visit. Summarize a voicemail. Extract line items from a photographed invoice. Classify an inbound message before routing it. None of that needs a frontier model. All of it is currently a per-token line on someone's invoice — and every one of those calls also ships customer data to a third party.

The pattern we build toward is boring and it works: route by difficulty, not by habit. Small, well-defined, high-volume tasks run on-device or on a small local model. The genuinely hard tail escalates to a hosted frontier model. We wrote up the same architecture at the server layer and why routing beats one big model — a phone TPU is just that idea arriving in your employees' pockets.

The catch is that on-device capability only helps if your software can use it. If your workflow is a browser tab talking to one vendor's API, a faster phone chip does nothing for you. If your workflow is an app you own with a pluggable inference layer, this hardware generation is a cost reduction you get for free on the next device refresh. That difference is an architecture decision you make once, early, and live with for years.

Sources: Google, 9to5Google.