Google’s Video Boost makes your Pixel footage look noticeably better, and it does it by uploading your clip to Google’s servers and grading it in the cloud. That is the tell I keep tripping over in 2026: a good chunk of what gets sold as on-device AI is the cloud wearing an on-device costume, and the seam between the two is exactly where the interesting questions live.
I run more than one of these phones as a daily driver, so I’m not reading this off a spec sheet. The marketing has converged into a single shiny word, AI, slapped on everything from your call screen to your camera shutter, and the useful work right now is separating the features that truly execute on the silicon in your hand from the ones quietly phoning home.
Start with the gate nobody puts on the keynote slide. Google’s big 2026 push, Gemini Intelligence, announced at The Android Show in May, has hard hardware requirements: 12 GB of RAM, a flagship-class chip, and Gemini Nano v3. That last one is the real wall. Nano v2, the version sitting on a Pixel 9 or a Galaxy S25, can do the lightweight stuff, smart replies and call summaries, but it physically cannot run the multi-step task automation and on-screen widget generation that define the new tier. So the Pixel 10, Galaxy S26, and OnePlus 15 are in, last year’s flagships are out, and the Galaxy Z Fold 8 is reportedly the first device to ship Gemini Intelligence publicly when it lands in July. There’s a delicious bit of irony lurking here too: leaks suggest Google’s own upcoming base Pixel 11 might ship with only 8 GB of RAM, which would leave Google selling a phone that can’t run the AI feature Google is building its whole pitch around. I’ll wait for the real spec before I dunk, but if it holds, that’s a self-own for the ages.
Apple, meanwhile, is the cautionary tale the whole industry should be staring at. We are almost two years past the WWDC 2024 reveal and Apple still hasn’t shipped everything it promised that day. I went into the Siri side of that gap after WWDC 2026, and the short version is that the demo was a check the engineering couldn’t cash on schedule. Apple Intelligence does have one philosophical advantage worth respecting: it runs on 8 GB where Google demands 12, and it leans on Private Cloud Compute for the heavy lifting in a way that’s at least honest about the handoff. Honest about the handoff is more than most of the Android camp manages.
Samsung went the opposite direction, flooding the zone. The Galaxy S26 series, out since February, piles on agentic features like Now Nudge, wraps it in a customized Snapdragon 8 Elite Gen 5 with reworked thermals, and even ships the world’s first built-in Privacy Display, which is a real hardware flex rather than another software checkbox (the side effect being a panel that’s measurably dimmer than the S25 Ultra, privacy mode on or off). Samsung’s smartest move isn’t a single feature, it’s aggression: pushing Galaxy AI down onto older flagships so the install base balloons. I sketched how that competitive ground shifted across Nothing, Pixel, and Xiaomi recently, and Samsung’s land-grab is the clearest play on the board.
The Pixel still owns the part that matters to me most, which is the camera brain and the assistant context. Magic Editor tapping individual objects to move or erase them, Camera Coach walking you toward a better frame, Call Screen 2.0 catching spam at better than nine-in-ten in live testing, that’s the stuff that earns its place. But notice how much of the marquee Pixel magic, Video Boost especially, routes back to Google’s datacenters. The Pixel is the best on-device AI phone and one of the best cloud-AI phones at the same time, and Google is careful not to make you think too hard about which is which.
Then there’s Nothing, and you know I have a soft spot here. Nothing’s whole design philosophy is transparency as strategy, and that ethos shows up in its camera stance: a true-to-life processing philosophy that deliberately refuses the over-sharpened, over-saturated look the big players lean on. In an era where the camera is mostly AI guessing what you wanted, Nothing choosing to guess less is its own kind of statement. The Phone 4a Pro holds its own on core camera work and undercuts the flagships hard on price, which is the Nothing trick: most of the experience, far less of the bill.
Step back and the honest pattern is the one the review crowd keeps landing on independently: in 2026, buyers finally care more about friction than fireworks. A full workday of battery without anxiety, a camera that handles a dim room without menu-diving, AI that quietly removes a step instead of adding a demo. Whether call screening or live translation is “the best AI feature” depends entirely on whether you get spam calls or talk across a language barrier, and if you don’t, half the marquee features are solving a problem you don’t have.
So where does that leave the actual question, is on-device AI useful yet? Partly. The truly local stuff, transcription, smart replies, the photo cleanup that doesn’t ship your shot to a server, has crossed from gimmick into daily utility, and the Nano v3 generation is a real step up rather than a version-number bump. If you want to push past what the system hands you, running your own local LLM on the phone is its own rabbit hole, and a more honest one about where the silicon actually is. But the headline features, the ones in the ads, still quietly lean on the cloud more often than the word on-device admits, and the hardware gating is getting stricter, not looser, just as the marketing insists the opposite. I’m not getting into the foldable AI angle here, the Fold 8 deserves its own look once it actually ships. What I’ll say is that I trust the feature I can use in airplane mode a lot more than the one that needs five bars to feel magic, and in 2026 that line is still doing most of the work.