← Longevity & Supplement Guides
6.4 to 7.8 Hours: The Sleep Window That Isn’t a Target
There is a particular arithmetic people do at eleven at night. Lights off at 11:20, alarm at 6:40, minus the twenty minutes of lying there — call it seven hours and change. Good enough. Then a headline arrives announcing that the sweet spot is 6.4 to 7.8 hours, and the arithmetic quietly becomes a verdict.
That headline has circulated for a fortnight, and it comes from a genuinely impressive paper. What the paper does not contain is a target. The gap between a finding and a target is most of what goes wrong when research meets a bedroom.
What half a million people and 23 clocks actually showed
The study, published in Nature in May 2026 by the MULTI Consortium with a team led at Columbia, is called a Sleep Chart. The design is elegant. Rather than test sleep against one summary measure of ageing, the authors built 23 separate biological ageing clocks — seven from organ MRI scans, eleven from plasma proteins, five from blood metabolites — and asked each one independently how it tracked with how long people said they slept.
The cohort is the UK Biobank, aged 37 to 84: 300,420 people in the normal six-to-eight-hour band, 16,872 short sleepers, 25,049 long sleepers. Across nine of the 23 clocks, the same shape appeared. Not a line. A U. Both ends of the sleep distribution looked biologically older than the middle.
The mortality numbers are the part that travelled. Against a six-to-eight-hour reference, sleeping under six hours carried a hazard ratio of 1.50 for all-cause death; over eight hours, 1.40. Across 726 disease endpoints tested, 153 associations survived correction. This is not a flimsy study. It is large, careful, well powered, and it earned its coverage.
It also, read closely, argues against the way it was reported.
Figure 1 · Where each clock says the bottom is
Sleep duration at which each biological ageing clock reached its minimum, UK Biobank (Nature, 2026)
What this study could not see (the authors' own list)
Sleep was remembered, not measured. A single touchscreen question; the authors list recall bias and misclassification first, and call for objective measurement.
The clock analysis is cross-sectional. One snapshot, so direction of effect cannot be established.
Long sleep as a marker of subclinical illness "cannot be fully excluded." Their words, in the limitations.
Circadian misalignment and sleep fragmentation were not directly assessed. Timing and continuity are absent from the analysis.
Predominantly European ancestry, which limits how far the numbers travel.
Protein and metabolite signals fluctuate with illness, medication and diet, so single snapshots can misclassify biology.
6.4 to 7.8 is not a window — it is a disagreement
Here is the sentence the headlines compressed. Across the nine clocks that showed the U, the sleep duration at which biological age bottomed out ranged from 6.5 to 7.8 hours in women and 6.4 to 7.7 in men. That range is not a tolerance band drawn around a single answer. It is the answer, and the answer is that the clocks disagree with each other.
Look at what disagrees. The brain's proteomic clock reached its minimum at 7.82 hours in women and 7.70 in men. The brain's MRI-based structural clock — same paper, same cohort, same organ — bottomed out at 6.48 and 6.42. Two ways of reading one brain, roughly eighty minutes apart.
So when someone quotes you an optimum of 6.4 to 7.8 hours, they are quoting the width of a scientific disagreement and presenting it as a prescription. Meanwhile fourteen of the twenty-three clocks showed no significant U at all. Whatever sleep duration is doing, it is not doing it everywhere in the body at once.
Which is the thing the optimisation industry keeps failing to absorb: your body does not have one age. It has a liver, a pancreas, an immune system and a skin, and none of them is obliged to agree with the others. A single biological-age number is a committee vote reported as a fact.
The measurement was a memory
Now the input. All half a million of those sleep durations came from one question on a touchscreen, asked once: about how many hours of sleep do you get in every 24 hours, naps included.
Consider what that question retrieves. Not last night — a habit, averaged by memory, in whole hours, by people who had come to a research clinic for other reasons. We know how well that works, because someone checked.
In 2008, researchers strapped wrist actigraphs to 669 middle-aged adults for three days at a time and compared the recordings with what those same people reported. Measured sleep averaged 6.0 hours. Reported sleep averaged 6.8. The correlation between them was 0.47 — real, but nowhere close to interchangeable. And the error was not random noise; it had a slope. For every extra hour of sleep people actually got, their self-report rose by only 34 minutes. Five-hour sleepers over-reported by about 1.2 hours. Seven-hour sleepers, by about 0.4.
If reported hours systematically compress real hours, a reported optimum of 7.8 corresponds to something meaningfully shorter in actual sleep — and the tails of the U, where over-reporting is worst, are exactly the parts everyone is quoting. None of this makes the U wrong. It makes the two decimal places furniture.
Figure 2 · Similar size, different animal
Excess all-cause mortality against a 6–8 h reference (reference = hazard ratio 1.00)
The two halves of the U are not the same finding
This is the part almost nobody reported, and it is the most interesting thing in the paper.
The authors ran a mediation analysis, asking whether the ageing clocks sat between sleep duration and later outcomes or off to one side. The two arms of the U behaved differently. For long sleep, the clocks did the carrying: the brain's MRI-based clock accounted for around 62% of the total association with late-life depression. For short sleep, the association was mostly direct, with only the adipose clock mediating.
Read plainly: long sleep looks like something that travels through an already-older body. Short sleep looks more like something that acts on one.
The authors are careful here, and their limitations section says the quiet part aloud — residual confounding and reverse causality, "particularly for long sleep as a marker of subclinical illness," cannot be fully excluded. Their Mendelian randomisation found no widespread evidence that disease drives sleep, but the depression-specific test showed pleiotropy (MR-Egger intercept p = 0.047, MR-PRESSO global p = 0.002) and the pleiotropy-aware estimators attenuated towards nothing. They call the causal question inconclusive. They are right to.
Long sleep has been the suspicious arm of this U for a long time. A 2018 meta-analysis pooled 137 prospective cohorts and 5,134,036 people: long sleep, risk ratio 1.39 for mortality, 1.26 for incident diabetes, 1.25 for cardiovascular disease. Consistent, enormous, replicated — and its authors closed by asking whether the relationship is causal or modifiable at all. Eight years on, that question is still open.
Which matters enormously for what you do on Sunday morning. If you sleep nine hours and still wake unrefreshed, this literature does not say "set an alarm." It says the nine hours may be information about your body rather than a habit in need of discipline. Cutting them short would edit the readout, not the thing being read.
The variable they left out is the one you can move
Tucked into the limitations is a line that quietly reframes the whole paper: circadian misalignment and sleep fragmentation were not directly assessed.
So this is a study about the one dimension of sleep people control least and misreport most — total duration — and it is silent on timing and continuity. Which is a shame, because timing is where the strongest evidence now sits.
In 2024, a team computed a Sleep Regularity Index from more than ten million hours of accelerometer data across 60,977 UK Biobank participants. Same biobank, but measured rather than remembered. The most regular fifth had a 20% to 48% lower risk of death over follow-up than the least regular fifth, and regularity outperformed duration as a predictor of all-cause mortality.
Bigger cohort for objective sleep, cleaner data, stronger signal — and no number with two decimal places, which may be exactly why it never became a headline. We went through it in detail in sleep regularity versus duration.
What the paper actually supports
I have spent five sections being difficult, so let me say what survives.
Both tails are real. They appeared across three independent measurement technologies — imaging, proteomics, metabolomics — which is far harder to produce by accident than one questionnaire correlation. Nine clocks built on different biology agreed on the shape. That is not noise.
The organ-level view is a genuine advance, and it is the finding I would keep. Sleep duration did not age everything uniformly: it registered in brain, immune, pulmonary, hepatic, skin, endocrine, adipose and pancreatic readings, and not in the rest. A body is a federation of systems ageing at their own rates, and this is the first sleep study at this scale to read it that way. It is also a useful corrective to the habit of compressing a whole night into one score.
And the direction of travel is not in dispute. Chronically short sleep is not good for people, and nothing here argues otherwise. The argument is narrower, and I think more useful: the paper found a shape, and the shape does not have a decimal point in it.
What to do with this on a Wednesday night
Five things, in descending order of confidence.
- Do not shorten your sleep to hit a number. Nobody in this study was assigned a sleep duration — not one person. Every figure in it is an association, and using an association as a set-point is how people end up sleeping worse in pursuit of a better reading.
- Watch your wake time, not your hours. It is the variable you actually control, it is the lever behind regularity, and it is observed rather than recalled. If you change one thing this week, shrink the gap between your earliest and latest wake time.
- Read a week, never a night. This study's own input was a blurred average, and a blurred average is closer to the truth than any single night. A seven-day trend is signal; Tuesday is weather. The same logic applies to what a weekend lie-in can and cannot repay.
- If you reliably sleep more than nine hours and still wake tired, take it seriously as information — not as laziness, and not as a discipline problem. In this literature long sleep behaves like a readout. That is a conversation to have with your doctor, not with an alarm clock.
- Stop collecting biological ages. If nine clocks in a single paper disagree by eighty minutes about one brain, a lone number claiming to price your entire body deserves scepticism rather than a subscription.
This is roughly why the Agen app shows sleep as a trend with a range rather than a nightly grade, and why the Band's job is to make consistency visible instead of scoring a night out of a hundred. Those are wellness estimates and trends to notice, not medical measurements or a diagnosis — but a trend you can act on beats a precise number you cannot. The boundaries are in what your wearable can and cannot measure.
The bottom line
A very good study found that both ends of the sleep distribution look biologically older than the middle, across nine independent clocks and three technologies. The same study found that those clocks disagree about where the middle sits by roughly eighty minutes, that its input was a remembered whole number, that the long arm of the U may be a symptom rather than a habit, and that the two variables most worth changing were never measured at all.
So: give yourself enough room for seven or eight hours, go to bed early enough to actually get it, hold your wake time steady, and then stop doing arithmetic at eleven at night. Use the numbers to correct a fantasy, not to replace the experience of waking up.
