Google unveils the Pixel 11 and its 'first 2nm chip' tonight. Nearly every number already circulating is a leak.

Google will walk on stage in New York this evening and unveil the Pixel 11. I am writing this a few hours before the keynote, which means I am writing it while the phone is, in the only sense that matters to me, still a rumour with a price tag. And yet the internet already knows its benchmark scores, its chip's manufacturing process, its camera specs, and — depending on which post you read — either two or three different prices for the same handset. That is the part worth stopping on. A device nobody outside Google has independently tested is already surrounded by numbers being repeated as fact. So before the confetti, let me do the boring thing and ask of each one: measured how, by whom, and compared to what?

To be clear about what is and isn't confirmed: Google has announced the event, teased the design, and told us the phones lean hard on Gemini. Everything else below — the chip, the scores, the prices, the on-device AI claims — comes from leaks, supply-chain reports and pre-release benchmark databases. I am not calling any of it wrong. I am pointing out that none of it has error bars yet, and that a launch is precisely the moment the gap between a claimed number and a checked one is easiest to hide.

The benchmark comes from firmware that isn't finished

Start with the number doing the most work in the reposts: a leaked Geekbench listing for the Pixel 11 Pro XL, its new Tensor G6 chip turning in roughly 2,112 on the single-core test and 5,196 on multi-core, with 16GB of RAM. It reads like a result. It is not one. It is a pre-production device running unreleased firmware, uploaded to a public database, screenshotted, and passed around. Benchmark databases are useful precisely because they are open, but that openness means what you are seeing is a snapshot of an unfinished phone, not a review-unit measurement under controlled conditions.

Now do the thing the reposts skipped and ask 'compared to what?' The Tensor G5 in last year's Pixel 10 Pro XL sits at roughly 1,876 single-core and 5,102 multi-core on the same test. So the leaked G6 numbers are about seven to thirteen percent faster on single-core and — read this slowly — essentially flat on multi-core. That is not the 'forty percent CPU leap' that has been floating around in the same breath. The forty-percent figure is a projection someone attached to the chip months ago; the leaked score is the closest thing we have to a measurement, and it quietly contradicts the projection. When the estimate and the data disagree, the field reposts the estimate, because it is the bigger number.

There is a further footnote that should make you slower still, not faster. An earlier leaked listing for the same chip, back in February, showed around 845 single-core and 2,657 multi-core — less than half of today's figures. Nobody thinks the chip doubled in performance in six months. What changed is the firmware maturing from 'barely boots' toward 'ready to ship.' That volatility is the whole point: a pre-release benchmark is a moving target, and the honest way to treat any single reading of a moving target is as one frame of a process, not the finish line. The GPU story is the one genuinely eye-catching leak — a compute figure reported as roughly eighty percent above the G5 — and it may well hold up. But 'may well hold up' is a sentence about the future, and I would like to run it before I believe it.

'2nm' is a name, not a measurement

The headline claim of the night will be the chip: the Tensor G6, codenamed Malibu, reported to be built on TSMC's 2-nanometre process — which, if it ships tonight, would make the Pixel 11 the first phone on the market with a 2nm chip, beating Apple's next iPhone to it by roughly a month. That 'first to market' line is true and almost entirely uninformative, and it is worth being precise about why.

'2nm' is a marketing name. It is not the size of anything on the chip. The 'nanometre' label stopped corresponding to a physical transistor dimension somewhere around 2009; since then it has been a generational badge that each foundry defines for itself. TSMC's '2nm' node — N2 — is characterised by a contacted gate pitch of roughly 45 nanometres and a tightest metal pitch near 20, and its actual novelty is architectural: it is TSMC's first gate-all-around transistor, a genuine engineering shift. None of that is captured by the number '2,' and none of it tells you how the phone in your pocket will behave. 'First 2nm phone' is a fact about a calendar and a supply contract. It is not a fact about performance, and it is definitely not the same sentence as 'the fastest phone,' however much the staging tonight will invite you to hear it that way.

'First 2nm phone' is a fact about a calendar and a supply contract. It is not a fact about performance — however much the staging tonight will invite you to hear it that way. — On the difference between being first and being fast

Perf or power — pick one

Here is what a new node actually buys, in TSMC's own published terms, because the vendor's own numbers are more honest than the marketing built on top of them. Moving from the previous generation to N2, TSMC quotes something like a ten-to-fifteen-percent speed improvement at the same power, or a twenty-five-to-thirty-percent power reduction at the same speed, plus around fifteen percent more transistors in the same area. Read the 'or.' Those are alternatives, not a bundle. A chip designer chooses where on that curve to sit — faster, or cooler, or denser — and cannot have all three at once. For a phone, the interesting choice is usually to spend most of the gain on power and heat rather than raw speed, because a cooler chip is one that can hold a high clock longer before it throttles, which is exactly the kind of improvement that never shows up cleanly in a thirty-second benchmark and only reveals itself in sustained use.

That, incidentally, is the charitable and probably correct reading of the flat multi-core leak: if the G6's gains went mostly into efficiency, a headline benchmark would look modest while the real-world experience — battery, thermals, how long the on-device AI can run before the phone gets hot — quietly improves. I would bet that is closer to the truth than 'forty percent faster.' But notice that this is me supplying the generous interpretation the marketing won't, because the marketing wants the big number and the efficiency story doesn't have one. The honest verdict is that a node this new is worth being calm about: N2 only reached volume production late last year, early trial yields were reported in the sixty-to-seventy-percent range, and Apple reportedly booked more than half of TSMC's initial 2nm output — which is to say the supply is tight, the process is young, and 'first' partly means 'first through a door that just opened.'

The on-device AI claims arrive with no test attached

The other pillar of tonight will be AI: 'Gemini Intelligence' branding, an expanded set of features that run on the phone itself rather than in Google's cloud — offline processing, screen awareness, multi-step tasks. This is a real and genuinely interesting direction. It is also the category where the numbers are thinnest. The most concrete technical claim in the leaks is that the new Android release uses aggressive four-bit quantization to shrink the on-device model's memory footprint by something like four to eight times versus a full-precision version. That is a plausible and specific-sounding figure, which is exactly why it deserves the question: shrink measured how, and at what cost to quality?

Quantization — storing a model's numbers at lower precision to make it smaller and faster — is a real and well-understood technique, and it always trades something. Push it far enough and the model gets measurably worse at the long or unusual tasks, in ways that a footprint number cannot see. 'Four-to-eight-times smaller' tells you about the storage. It tells you nothing about whether the compressed model still answers as well, and 'runs on the phone' is a claim about where the computation happens, not how good it is. Notice, too, what is missing: not one of the leaks I can find carries a straight, comparable measure of the on-device AI's throughput, nor even a confirmed name for which Gemini model actually runs locally. There is branding. There is no benchmark. When a capability is announced entirely in adjectives, the adjectives are the product.

'Compared to what' bites hardest here. On-device AI is only meaningful against a baseline — against last year's phone, against the same task run in the cloud, against a rival's silicon. Tonight will offer demos, and demos are the opposite of error bars: a demo is a single cherry-picked run, staged by the party with the most to gain, on hardware you cannot inspect. I will be delighted to be impressed. I would just like to be impressed by something I can measure.

The price is three different numbers

If you want a clean demonstration that these figures are leaks rather than facts, look at the prices, because facts do not usually come in threes. The base Pixel 11 is widely reported to be moving to around 899 dollars, up roughly a hundred from last year, with the cheapest storage tier quietly dropped — a price rise wearing the costume of a spec bump. The Pro is pegged near 1,099 and the Fold near 1,899. And the Pro XL is listed at 1,299 in one well-sourced leak and 1,199 in another. Both cannot be right. That fifty-fifty split is not a scandal; it is a thermometer. It tells you the reporting is genuinely ahead of the announcement, that these numbers are being assembled from supply chains and retailer databases and not read off a Google slide — which is the same reason to hold every other figure in this article a little more loosely than the confidence around it suggests.

What tonight settles, and what it doesn't

So let me separate the significant from the impressive, which is the only distinction that survives a launch. Significant: Google appears to have gotten a 2nm, gate-all-around chip into a shipping phone first, it has switched its modem supplier, and it is pushing more AI onto the device itself. Those are real strategic facts, and they matter — they tell you where Google is spending and what it is betting on. Impressive is a different word, and it is a measured one. Whether the G6 is meaningfully faster, whether the efficiency gain is real in your hand, whether the on-device AI is good rather than merely present — none of that is settled by anything circulating today, and only some of it will be settled by the keynote, because a keynote is an advertisement with better production values.

Here is the checklist I will be holding against tonight's claims, and I would suggest you hold it too:

• Sustained performance, not the opening-second peak: how the chip behaves after ten minutes of load, which is where an efficiency-first node either pays off or doesn't.
• Battery and thermals measured by someone who didn't build the phone, on a retail unit, not a demo device.
• An on-device AI number with a baseline attached — versus the cloud, versus last year, versus a rival — instead of an adjective.
• Whether the leaked benchmark rises to meet the projections once final firmware ships, or whether the projections quietly climb down to meet the benchmark.
• The price you actually pay after the storage tier shuffle, which is the one number tonight will state plainly and the one the leaks can't agree on.

None of this is Pixel-pessimism. Google has built genuinely good phones, and the Tensor line has quietly become the most interesting bet in the category precisely because it is willing to spend silicon on AI rather than on chasing benchmark crowns. I hope the G6 is excellent. But 'I hope' is not 'I measured,' and the distance between them is where every launch does its most persuasive work. Tonight you will be handed a stack of numbers polished to a shine. The useful thing you can do with them is the same thing you can do with any number the source chose for you: write them down, and wait for the ones that survive being tested by someone who wasn't on the stage. If the chart only goes up, look harder. It goes double on the night the chart is being drawn for you.