How accurate is Apple Watch VO2max? And Garmin?

Apple Watch VO2max is a decent estimate that tends to read low if you're fit. In the most careful independent study so far, it came in 6.07 ml/kg/min under the lab value on average, and the 95% limits of agreement for one person ran from about 6 above the lab value to 18 below it.¹ Garmin looks a bit tighter on paper, around 5 to 7% average error, but nobody has tested the two side by side, so treat both as a range, never a single number.

What Apple says, and who it tested#

Apple's own validation is in a 2021 white paper.² The headline is good: in 221 people held out from the algorithm's design, the watch was off by 1.4 ml/kg/min on average, with a standard deviation of 4.7, and readings 28 days apart agreed well (ICC 0.86).

Look at who those 221 people were, though. Their average age was 55, their average reference VO2max was 29.7, and about one in ten had coronary artery disease. That's a fine group if you're building a health feature. It isn't you, if you're reading a blog about lifting and running. The reference test also wasn't a true max test: Apple projected each person's submaximal heart rate and oxygen data out to an age-predicted max heart rate. So the watch was graded against another estimate.

The paper is also honest about what drags the number down: wrong age, sex or weight in the Health app, carrying a heavy load, soft ground like sand, and anything that pushes heart rate up, including heat, dehydration, caffeine and altitude.²

What feeds the estimate, and what doesn't#

This part matters more for hybrid athletes than for runners. Apple's estimate only comes from outdoor walks, runs or hikes on fairly flat ground (under 5% grade), with decent GPS and a heart rate that climbs at least about 30% of the way from resting to max.² Your first qualifying workout won't produce a number, and in Apple's data about 79% of outdoor walks and runs longer than 5.75 minutes produced one.

So the sled push, the SkiErg, the rower, the wall balls, a treadmill session, a spin class and every lifting day add nothing. If most of your conditioning happens indoors, your watch might be grading your engine off three easy outdoor jogs a month. That's a thin sample for a number you're going to compare against a chart.

Garmin's method has the same dependency on heart rate and speed from runs (and power for cycling), so the same logic applies: the estimate only knows the sessions it can read.³

What independent labs found#

Two independent studies put Apple Watch next to a gas analyser, and both found the same direction.

Lambe and colleagues tested 30 adults in 2025, mostly on a Series 9 or Ultra 2 running watchOS 10 or later, and compared each person's watch reading with a maximal treadmill test.¹ Two people didn't reach a valid max, so 28 were analysed. The watch underestimated by 6.07 ml/kg/min on average (95% CI 3.77 to 8.38), the mean absolute error was 6.92, and the mean absolute percentage error was 13.31%. On someone whose lab value is 50, that's an average miss of about 6.7 ml/kg/min, which on the US tables is most of the way from the median (48.0) to the 75th percentile (55.2) for a man in his 20s.⁴ The catch the authors point out themselves: 21 of their 30 participants were classed Superior or Excellent on the FRIEND fitness standards. That's a weakness for generalising to everyone. For a hybrid athlete it's the most relevant sample there is.

Caserman's group tested 19 people with a Series 7 in 2024.⁵ The lab average was 45.88 and the watch average 41.37, a mean absolute percentage error of 15.79%. Their lab test was on a bike while the watch learns from running and walking, which usually makes the lab number lower than a treadmill test would,¹ so if anything that setup flatters the watch. Their conclusion is the useful bit: the watch overestimated people with poor fitness and underestimated people with excellent fitness.

Watch minus lab VO2max: average and 95% limits

	Low (ml/kg/min)	High (ml/kg/min)	Mark (ml/kg/min)
Apple Watch, Lambe 2025	-18.26	6.11	-6.07
Exercise-based wearables, meta-analysis	-9.92	9.74	-0.09
Resting-based wearables, meta-analysis	-13.07	17.41	2.17
Apple's own validation, 2021	-7.8	10.6	1.4

Mark is the mean bias (watch minus lab), bar is the 95% limits of agreement. Lambe 2025 (Apple Watch, n = 28); Molina-Garcia 2022 meta-analysis (14 studies, no Apple Watch); Apple 2021 white paper limits computed by us as mean plus or minus 1.96 SD, against a submaximal reference.

What about Garmin#

Garmin's VO2max comes from an algorithm developed by Firstbeat, which reads the relationship between your heart rate and your speed across a run and keeps only the segments that pass its error checks.⁶ Firstbeat's 2017 white paper reports about 5% mean absolute error for running, from 2,690 runs by 79 runners who were lab-tested four times while preparing for a marathon, with most errors under 3.5 ml/kg/min.³ That's a manufacturer document, not a peer-reviewed study.

The same paper flags the weak spot: the method needs your max heart rate, and it guesses it from age if you don't set it. Set 15 beats too low, error goes to about 9%. Set 15 too high, about 7%.³ If you've never told your watch your real max, fix that first.

A 2025 independent study tested the Garmin fenix 6 on 19 people with an average lab VO2max of 48.9.⁷ Mean absolute percentage error was 7.05% and the concordance coefficient 0.73, which passed the authors' own validity bar (under 10% and above 0.7). It failed their equivalence test, though, so "close on average" is the fair summary.

The main meta-analysis, INTERLIVE 2022, pooled 14 studies on Polar, Garmin and Fitbit. No Apple Watch study existed yet.⁶ Devices that estimate from exercise (heart rate plus pace) had almost no average bias, minus 0.09 ml/kg/min, but 95% limits of agreement of about plus or minus 10. Devices that estimate from resting data did worse on both counts. The authors rated 11 of the 14 studies at high risk of bias and concluded that the error for one individual is large. The paper also picked up a correction notice in October 2026, which wasn't accessible when this was written, so the numbers here are the original ones.

So which watch is more accurate?#

The data can't say. The Garmin and Apple studies used different people, different watch models, different lab protocols and different error statistics. The larger independent Apple study (13.31% error) used a very fit group, and both Apple studies suggest fitness is exactly where it drifts. Garmin's numbers come mostly from runners and from Firstbeat's own data.

What you can say: both are exercise-based estimates, and that whole class sits at roughly 10 ml/kg/min either side of the lab for one person.⁶ On the US reference tables, that's the difference between the 25th and roughly the 63rd percentile for a man in his 30s.⁴ Use the VO2max by age chart with that in mind.

My own watch, against my own races#

Here's what this looks like on one athlete. From April to the end of May 2026 my Garmin logged a VO2max on 60 days. The values ran from 50.1 to 59.9, and they didn't drift, they flipped: 34 days sat between 50.1 and 53.9, and 26 days between 59.1 and 59.9. The log doesn't say why it jumped between the two.

In March 2026 I ran a half marathon I titled "6 MIN PR" on Strava: 21.32 km on GPS in 92 minutes. Run through Daniels and Gilbert's VDOT formula, that's about 49.7 on the official half distance, 50.3 if you trust the GPS.

And on 22 October 2026, the app's estimate from the heart rate and pace of 10 recent steady runs was 47, plus or minus 5.

Three readings of one engine

	Low (ml/kg/min)	High (ml/kg/min)	Mark (ml/kg/min)
Garmin daily readings, Apr to May 2026	50.1	59.9	53.5
Half marathon VDOT, Mar 2026	44.2	55.2	49.7
hybridAF estimate, Oct 2026	42	52	47

Ken's data, read 2026-10-22 from his hybridAF log. Garmin: daily values Apr to May 2026, mark at the median. Half marathon: VDOT from 92 minutes for 21.1 km, bar is plus or minus 5.5, the error hybridAF assigns an unconfirmed race (src/lib/vo2max-estimator.ts). App: heart-rate estimate, Oct 2026, plus or minus its stated range.

Different months, so some of the gap could be real fitness change. But the lesson stands: the watch's own spread over two months was almost 10 ml/kg/min, about the same width as the meta-analysis limits. The half marathon, the only one of the three that comes from a race actually run, lands in the middle of everything.

How to sanity-check your reading#

Use a race. A recent all-out effort between about 5K and a half marathon gives you a VO2max estimate that reflects what you actually ran, with the pace and the time built in. For scale, a 20:00 5K works out to a VDOT of about 49.8, and a 25:00 5K about 38.3. If your watch says 55 and your best recent all-out 5K is 25 minutes, believe the 5K.

A few more things help. Set your real max heart rate, because it moves Garmin's error by several percent.³ Keep your Health profile current, since age, sex and weight all feed Apple's estimate.² Compare like with like: same watch, same kind of run, similar weather. A hot week can pull the number down without your engine changing.²

And watch the trend, not the day. In Apple's data the median spread of one person's estimates was a standard deviation of 1.2 ml/kg/min, so a change that small is noise.²

How hybridAF handles a watch reading#

The app counts an Apple Watch reading as one input among several, with an error of about 8.7 ml/kg/min built from the Lambe study's bias and limits. It doesn't subtract the 6-point bias, it just trusts the reading less. A hard race confirmed by heart rate is weighted with an error of 4.5, so it counts for roughly three to four times as much. If you want to see what your own races say, the VO2max calculator runs the same equations.

Whatever you use, the number is a ceiling estimate, and the ceiling isn't the whole engine. For HYROX especially, what you can hold for an hour matters more, which is the subject of lactate threshold vs VO2max. The rest of the engine reading list is on the engine hub.