The Libre and Dexcom CGMs shared an arm. They didn’t share a GMI.

At one point in September my Dexcom G7 read 180 mg/dL (10.0 mmol/L) while the Libre 2 Plus on the same arm read about 140 (7.8). Neither was obviously broken. Both were tracking the same rise, a few minutes apart, and they simply disagreed on where it was.

Over eleven weeks that kind of disagreement added up to two different versions of me. From the G7 my estimated A1C was 6.59%. From the Libre it was 6.11%. The G7 said I spent 3% of the time below 70 mg/dL (3.9 mmol/L); the Libre said 9%. A clinic looking at the first would see someone comfortably inside every consensus target, and one looking at the second would want to talk about hypos.

This article is about where that gap came from and what it does to the numbers a clinic acts on, which turned out to depend on which pair of sensors happened to be on.

How I compared them

From 11 August to 27 October 2026 I wore a run of eight G7 sensors alongside six Libre sensors: two Libre 3+ and four Libre 2 Plus, all on the upper left arm. The G7 was read through xDrip and the Libres through Juggluco, so what I compared was the whole chain from sensor to stored number, and neither manufacturer’s own app is in it. For nine days around the start of September two Libres were on at once, read by different phones.

Readings from the two systems were paired when they landed within 40 seconds of each other, which gave just under 17,000 matched five-minute points over 67 days. Thirteen different G7 and Libre combinations were on together at one time or another, and I’ll call each of those a pairing.

There was no laboratory reference running alongside, so nothing here says which sensor was accurate. What it can say is how far apart they were over time, and what the gap did to the numbers people actually use.

How far apart they read

Across the whole period, the Libre read 19.9 mg/dL below the G7 (95% interval 16.6 to 23.4), about 15% of the G7’s average. The gap was not fixed. It grew as glucose rose: on the first Libre 2 Plus a G7 at 180 corresponded to a Libre at 140, and on the last one to a Libre at 155.

Figure 1 shows each Libre sensor’s average gap from the G7, in the order I wore them, with its interval. Five of the six read below the G7, by 16 to 34 mg/dL. One, a Libre 3+, read 1 mg/dL above it.

It is tempting to read Figure 1 as a ranking of Libre sensors, and I did at first. Two Libre 2 Plus sensors worn one after the other read 34 and 23 mg/dL below the G7, which looks like two sensors of the same model disagreeing by 11. But each Libre was measured against whichever G7s were on at the time, and the G7s turned out to move the gap as well.

What happens when you change a sensor

Five times a G7 was replaced while the same Libre carried on. At four of them the Libre’s reading relative to the G7 dropped by 13 to 46 mg/dL, and that held when I compared only stretches of steady glucose and when I allowed for glucose running higher or lower across the change. At the fifth the shift was about 3 mg/dL and not consistent. The one time a Libre was replaced while the same G7 carried on, the gap moved by 26 mg/dL.

This demonstrates there’s perhaps more intersensor variability that we tend to talk about. On its own it can’t tell you which sensor moved. Three of those four big shifts also happened while the Libre that carried on was within a day of going in or within two days of coming off, which is when a sensor is least settled. The cleanest one is the G7 change on 6 October, against a Libre in the middle of its wear: the gap jumped by 28 mg/dL within six hours and stayed there.

The important point in the above is that neither the Libre series nor the Dexcom remained consistent across sensors.

What it does to your estimated A1C

Most CGM reports now include a glucose management indicator, GMI, which is the A1C you would expect from your average glucose: 3.31 plus 0.02392 times the mean in mg/dL. Every 10 mg/dL of average glucose is about a quarter of a point.

The right-hand axis of Figure 1 turns each sensor’s gap into GMI. On the whole dataset the difference was 0.48 points (0.40 to 0.56). Sensor by sensor it ran from nothing, on the Libre 3+ that agreed with the G7, to 0.8 of a point on the Libre 2 Plus that read lowest.

For scale, a laboratory A1C method has to land within 5% of the reference to be certified, and proficiency testing allows 6%. Half a point at 6.6% is about 7%. The key point is that the gap between two popular sensors on one arm was bigger than a lab test is allowed to be wrong by.

None of this changes anyone’s actual A1C, which is measured from blood. What changes is the estimate printed beside it. Move from a G7 to a Libre and your GMI could drop by half a point with nothing about your glucose having changed. Change a single sensor, of either kind, and it can shift by several tenths.

Against my laboratory A1C

I had a laboratory HbA1c on 29 September: 5.5%. Over the 90 days before it, the G7 gave a GMI of 6.38%, or 6.47% if the recent weeks are weighted more heavily, as red cells weight them. The Libre, over the shorter record I had for it then, gave 6.10 to 6.14%.

So both sat above my lab result, the G7 by about 0.9 of a point and the Libre by about 0.6. It would be easy to read that as the Libre being more accurate. It doesn’t show that. GMI and laboratory A1C differed by half a point or more in 28% of the people the GMI equation was built from, and in half of 641 clinic visits in a later study, because red cell lifespan and iron status shift A1C at the same average glucose. My lab result is a gap that a clinic would see anyway, and it’s the same size as the gap a change of sensor makes. It can’t tell you which sensor is right.

Against fingerpricks

I also took eight fingerpricks over the period. Four of them were while glucose was moving fast, and on those the sensors and the meter disagreed significantly, which is what a few minutes of delay does on a steep fall.

The other four were taken while glucose was steady. On those the G7 averaged within a fraction of a mg/dL of the meter (a mean difference of 0.1 mg/dL, and about 9% reading by reading), and the Libre averaged about 24 mg/dL below it on the three where a Libre was on. The readings were taken as soft calibrations and used a meter that is itself allowed 15%. Three fingerpricks in all of the data mean nothing. The only thing to take is that it fitted the 20 mg/dL gap between the sensors, and the meter sided with the G7 on those occasions.

Lows and the targets

Figure 2 is where the gap matters most. A sensor reading 15 to 25 mg/dL lower crosses 70 whenever the other one is in the 80s, which is where my G7 often was. On every Libre sensor except the Libre 3+ that agreed with the G7, the Libre showed more than the consensus 4% of time below 70, and on the last two the G7 showed under 2%.

Clinics look at 14 days, so I split the record into fortnights from 11 August. The consensus wants about 70% of the readings over those 14 days before the numbers are interpreted, and four fortnights had that much with both sensors on. In three of the four the two sensors gave opposite answers on the time-below-range target:

Fortnight	G7, time below 70	Libre, time below 70
8 to 21 September	3.0%	5.6%
22 September to 5 October	2.1%	7.0%
6 to 19 October	0.8%	6.3%

In the fourth, the first fortnight, neither met it (6.3% on the G7 and 16.9% on the Libre). Working the same fortnights out from all of each sensor’s own readings, including the times the other sensor had a gap, gave the same answers. Both sensors gave the same answer on time in range in every fortnight.

How much of the difference in lows is just the level? Within each pairing I scaled the Libre readings up so their average matched the G7’s, and looked at time below 70 again. Pooled, the gap fell from 6.4 percentage points to 0.6. In six of the seven pairings that lasted four days or more it ended up within 1.8 points. That doesn’t make it trivial. Which way the lows go depends on a level difference that wouldn’t stay put, and that level difference is the thing that moved at every sensor change.

Correcting for it

The obvious idea is to measure the gap once and adjust. A correction fitted on my earlier Libre sensors predicted the last one almost exactly. Fitted on five sensors and tested on the sixth in turn, it held to within 4 mg/dL for four of them and was out by 11% and 21% for the other two, in opposite directions. The straight-line version fixed the average better but pushed the Libre’s time below 70 under the G7’s on every sensor it was tested on. So aligning a correction on each sensor proved tricky, which is worth considering if you’re adding calibration to Libre sensors, which famously are non-calibratable. You mustn’t carry the calibration across sensors.

What should we as users do with this?

When you share CGM numbers, with a clinic or anyone else, say which sensor and which app produced them, the same way a lab report states its method. An A1C estimate or a time-below-range figure from a G7 and one from a Libre aren’t the same measurement.

Expect a step in GMI and time below range when you change sensor model, and sometimes when you just change sensor. If your numbers jump the week you switched, look at the switch before you look at yourself.

If you loop, a sensor reading low makes the loop dose as if you were lower. That leans safer on lows and slower on highs. My earlier results on loop cadence used a Libre for dosing, which is one reason I read their time-below-range figures as comparisons between settings rather than as absolute levels.

What this can’t tell you

This is one person, one arm, six Libre sensors and eight G7s, each read through third-party software rather than the manufacturer’s app. There was no laboratory reference glucose, so it measures disagreement, not error. No two sensors of the same model were worn together, so I can’t split the movement between the Libres and the G7s. A study of several sensors in more people would be needed to separate a difference between the manufacturers from ordinary variation between sensors.

The full analysis, with intervals for everything above, is in the measurement preprint that will be published on SSRN, “Acquisition-chain dependence of CGM metrics”. How far the Libre trails the G7 in time is the subject of a separate post.

In summary…

If you ever wondered why your Hba1C and your predicted GMI from your CGM didn’t match, this might help explain why, especially if you are a Libre user. The bigger challenge here is that with this level of variation across two sensor families, how can Hba1C be replaced by GMI or even the updated version? It’s not possible to compare the two different values when there’s no frame of reference against which each sensor family can be quoted.

Within research circles there is an ongoing argument that Hba1C is now a redundant measure, for a lot of reasons, and that Time In Range and eGMI should replace it. While the logic for that seems to make sense, this n=1 dataset suggests otherwise.

It reminds me of George Orwell. “All GMIs are equal, but some GMIs are more equal than others.”