What do biological age tests actually measure
Longetic keeps the report a longevity test sends you next to your ordinary lab results, and it does not compute a biological age, a health score or a risk score of its own. The tests sold as biological age fall into three families, and they were trained on three different targets: your chronological age, your risk of dying, and how fast your body is changing. A number from one family is not comparable to a number from another.
The confusion is built into the product name. "Biological age" sounds like a measurement of a thing, the way body temperature is. It is the output of a statistical model, and what the model was trained to predict determines what its output means.
| Family | Founding paper | Trained to predict | What the paper reported |
|---|---|---|---|
| First generation, Horvath | Horvath, Genome Biology 2013 | chronological age, across 51 tissue types | test correlation 0.96, error 3.6 years |
| First generation, Hannum | Hannum, Molecular Cell 2013 | chronological age from whole blood | a blood-specific age estimate |
| Second generation, PhenoAge | Levine, Aging 2018 | a composite of 9 clinical markers plus age | one extra year of PhenoAge, 4.5% higher all-cause mortality risk |
| Second generation, GrimAge | Lu, Aging 2019 | time to death, via surrogates for 7 plasma proteins and smoking | hazard ratio 1.10 per year for all-cause mortality |
| Pace of ageing, DunedinPACE | Belsky, eLife 2022 | 20-year rate of decline across 19 organ systems | correlation 0.78 with the cohort's measured pace, mortality hazard ratio 1.65 |
| Organ-specific proteomic clocks | Oh, Nature 2023 | organ-level age from plasma proteins | accelerated heart ageing, hazard ratio 2.37 for heart failure |
A first-generation clock that says you are 44 when you are 47 has told you something about how well it guesses chronological age. That is what it was built to do. It has not told you that you will outlive your birth certificate.
How much does the number move between two runs of the same blood
Enough to change the headline. When researchers ran 36 whole-blood samples in technical duplicate, PhenoAge differed between the two runs by a median of 2.4 years and a maximum of 8.6 years, on the same blood, in the same laboratory. GrimAge, built differently, differed by a median of 0.9 years.
The Higgins-Chen analysis in Nature Aging put actual reliability coefficients on the standard clocks.
| Clock | Intraclass correlation | Median deviation between duplicates | Maximum |
|---|---|---|---|
| Horvath, raw score | 0.945 | 2.1 years | 5.4 years |
| PhenoAge | within the 0.9 range | 2.4 years | 8.6 years |
| GrimAge, raw score | 0.989 | 0.9 years | 2.4 years |
| Principal-component versions of all six clocks | above 0.99 | most replicates within 1.5 years | 1.5 years |
| DunedinPACE | 0.96 | not reported in these terms | not reported |
The authors summarised the problem in one line: "Technical noise produces deviations up to 9 years between replicates for six prominent epigenetic clocks."
A 2025 follow-up found something less comfortable. Technical reliability, how closely two runs of the same sample agree, does not predict biological reliability, how well the test tracks real change in a person over time. Most clocks landed at an intraclass correlation of roughly 0.4 to 0.7 biologically despite exceeding 0.9 technically, and the correlation between the two measures across clocks was 0.0168, which is to say none at all.
Do two different tests agree about the same person
Often they do not, and the people who built these clocks say so first. Steve Horvath, author of the first pan-tissue clock, told The Scientist: "It's not yet possible that you can go to different validated labs and get the same readout." Different training targets, different sample types and ordinary measurement noise all pull the numbers apart.
He has the experience to back it. Horvath took a GrimAge test twice, roughly six months apart, and the second result came back four years older than the first. Morgan Levine, who built PhenoAge, told MIT Technology Review that results "can wind up being off by decades" and that the clocks "simply aren't accurate enough to count on in most cases".
A recent worked example: one researcher tested herself with three commercial products and got 33.7 years from one, 45.1 from another and 49.8 from a third, a spread of roughly 16 years for one person. Two of the three came from the same vendor, using different algorithms, which is arguably worse than a disagreement between competitors.
None of this means the underlying science is wrong. It means the number on the certificate carries an error bar that the certificate does not print.
Are biological age tests regulated
No biological age test is cleared or approved by the US Food and Drug Administration. They are sold as laboratory developed tests or as wellness products, and in the United States that route became less regulated rather than more during 2025, after a federal court vacated the rule that would have brought such tests under device regulation.
The FDA's 2024 rule that would have brought laboratory developed tests under medical-device regulation was vacated in full by the US District Court for the Eastern District of Texas on 31 March 2025, and the agency did not appeal. Laboratory developed tests remain under CLIA laboratory certification rather than FDA premarket review. In the European Union, in-house tests used inside a health institution are exempt from full CE marking under Article 5(5) of the In Vitro Diagnostic Regulation, but only where the institution meets the general safety and performance requirements and documents why no CE-marked equivalent exists, which is a narrower door than the US one.
The scientists building these measures are the most cautious voices in the room. The Biomarkers of Aging Consortium, writing in Cell, stated it plainly: "There are not yet any formally validated or likely surrogate endpoint biomarkers of aging, in part because they have only recently been described." The US National Institute on Aging writes about organ-level age tools in the conditional: if the new measures can accurately predict future disability and mortality, they could help guide lifestyle changes.
What do these tests cost
Published prices run from $249 to $499 for a single kit, and several vendors bundle the result into an annual membership instead of selling it outright. Two of the eight products in the table below did not show a price on their own page when we checked, and we marked those cells as not verified rather than filling them from a reseller.
| Product | Price as published | Sample | What it uses |
|---|---|---|---|
| TruDiagnostic TruAge Complete | $499 | finger-prick blood spot | OMICmAge, SymphonyAge and DunedinPACE |
| NOVOS Age | $349, or $299 a kit for two or more | finger-prick blood spot | DunedinPACE plus an organ-age algorithm |
| Elysium Index | $299 | saliva | a methylation array, reporting 9 system ages |
| Tally Health TallyAge | $249 | cheek swab | an epigenetic clock trained on cheek-swab data |
| Function Health | $365 a year membership | blood draw | biological age as one output of a large panel |
| InsideTracker InnerAge | $99 add-on, membership $149 a year | blood draw | blood biomarkers |
| GlycanAge | not verified, the vendor's price page did not load for us | finger-prick blood spot | IgG glycosylation |
| Thorne Biological Age | not verified, the vendor's price page did not load for us | blood draw at a partner laboratory | the Klemera-Doubal algorithm on blood markers |
One vendor claim worth reading carefully: InsideTracker states that "60% of customers reduce their InnerAge on their follow-up analysis". That is a marketing figure from the seller, not a published trial result, and with test-retest variation in the range described above, a fall on a second test is not by itself evidence that anything changed.
What actually predicts how well you age
The measures with the strongest published evidence are unglamorous and mostly free: blood pressure, lipids, glucose, kidney function, how fast you walk, and how hard you can grip. Every one of them appears in a guideline or a large cohort study, which is more than any commercial biological age score can say today.
| Measure | What the evidence shows | Source |
|---|---|---|
| Gait speed | each 0.1 m/s faster, hazard ratio 0.88 for survival, pooled across 9 cohorts and 34,485 adults 65+ | Studenski, JAMA 2011 |
| Grip strength | each 5 kg lower, hazard ratio 1.16 for all-cause mortality across 139,691 people in 17 countries, a stronger predictor of death than systolic blood pressure | Leong, Lancet 2015 |
| Life's Essential 8 | eight components scored 0 to 100: diet, physical activity, nicotine, sleep, body mass index, non-HDL cholesterol, blood glucose, blood pressure | AHA, Circulation 2022 |
| Cardiorespiratory fitness | argued as a clinical vital sign in an AHA scientific statement | Ross, Circulation 2016 |
| ApoB and Lp(a) | lipid targets and risk classification | ESC/EAS dyslipidaemia guidelines |
| eGFR and uACR | the kidney risk grid | KDIGO 2024 |
The scoring in Life's Essential 8 is worth knowing because it is public and you can read your own components off an ordinary report: 100 points for blood pressure under 120 over 80, 100 points for non-HDL cholesterol under 130 mg/dL, which is under 3.4 mmol/L, 100 points for seven to under nine hours of sleep, 100 points for at least 150 minutes of moderate activity a week.
That is not an argument that epigenetic research is pointless. It is an argument about what is decision-grade today.
What does the evidence in favour look like
The case for the clocks is real and getting stronger, and it rests on three findings in particular: epigenetic age acceleration predicts mortality after adjusting for chronological age, a pace-of-ageing measure moved in a randomised caloric restriction trial, and organ-specific proteomic clocks replicated across UK, Chinese and US cohorts. None of that makes a single consumer result reliable.
First, epigenetic age acceleration predicts mortality after adjusting for chronological age. A 2025 analysis reported a hazard ratio of 1.51 for GrimAge acceleration for overall mortality after adjusting for age, sociodemographics, behaviours and comorbidities, though the association did not hold consistently in Hispanic participants.
Second, a pace-of-ageing measure moved in a randomised trial. In the CALERIE-2 trial, two years of roughly 12% caloric restriction slowed DunedinPACE with effect sizes of -0.29 at 12 months and -0.25 at 24 months, both statistically significant, equivalent to a 2% to 3% reduction in the pace of ageing. The same trial found no significant effect on PhenoAge or GrimAge, which tells you something about which family responds to what.
Third, organ-specific proteomic clocks replicated across populations. A 2026 Nature Aging paper reported organ clocks generalising across UK, Chinese and US cohorts with correlations of 0.93 to 0.98, and predicting organ-specific disease, with a hazard ratio of 1.88 for dementia from brain ageing and 1.78 for chronic kidney disease from kidney ageing.
The honest summary is that this is a research field producing results, sold as a consumer product before the measurement error has been solved.
What to do with a result you already have
Longetic keeps the report and the ordinary markers in one searchable history, so a second test a year from now has something to sit next to. Keep the underlying data, not just the headline number. Every one of these tests also produces or requires real measurements, and those are the parts that stay meaningful when the algorithm behind the headline gets retrained.
Three practical things follow from the reliability numbers. A single result is a point estimate with several years of noise around it, so a change smaller than that is not a change. Comparing a result from one vendor with a result from another is not a comparison. And if you retest, using the same vendor, the same sample type and the same algorithm version is the only way the second number relates to the first.
Tools that store this vary in what they will take. Guava Health connects to US patient portals. KeepMD holds records, labs and visits in one timeline. Longetic reads a lab report you upload as a PDF or a photo, from any laboratory, charts each biomarker across every report, and keeps medications, diagnoses and visits in the same history. As a disclosure, Longetic is our product, so weigh that mention accordingly. It does not calculate a biological age, a health score or a risk score, it is iPhone only, and it does not tell you what any value means.
