CGM Guide
Which CGM is the most accurate? Wrong question.
Every CGM manufacturer wants you asking that question, because every one of them has an answer ready that makes their own device look best. Give me a few minutes before you look at a single spec sheet, because the question underneath it is more interesting, and more useful, than “which one wins”.
Ask Grace
Want the study-design score or the accuracy figures for a specific device explained in your own situation? Ask Grace.
First, know which kind you are holding
Every CGM on the market falls into one of two categories, and it changes what you are allowed to do with the number on screen.
Adjunctive
Designed to sit alongside finger pricks, not replace them. Every dosing decision, every correction, still needs a blood glucose confirmation. The CGM is context and trend, not the number you act on.
Non-adjunctive
Licensed to replace finger-stick blood glucose for treatment decisions, including insulin dosing. This is what almost everyone reading this guide is using or choosing between. Everything below is written for this category.
That single word on the licence, adjunctive or non-adjunctive, matters more than any spec on the box. It is the first thing worth checking on any device you are considering.
The question that actually matters
Here it is: accurate enough for what? A sensor that is close enough to trust for a bedtime glance is not automatically close enough to trust for a correction dose. Those are different jobs. Marketing material almost never tells you which one it is talking about.
So before I show you a single device, I want to walk you through how these numbers get made. A few years ago I sat down with the DSN Forum UK, the specialist nurses who actually help people choose a sensor day to day, and we built a filter together. Not a features checklist. A test of whether the study behind an accuracy claim is even worth reading.
Five questions, before a single number
A CGM earns one point for each of these. Five questions, five points, and a device needs all five before I will let its accuracy figures anywhere near this page:
- Is the evidence public? Peer-reviewed publication, or the data is at least available through a regulator (the FDA device database, or EUDAMED). Not “data on file”, not a number in a press release.
- Is there enough of it? Roughly fifty or more people wearing the sensor with insulin on board, generating thousands of paired readings against a real reference, not a handful of volunteers on a good day.
- Did they actually eat and dose? A meal and insulin challenge, so the sensor was tested against a real rise, not just a quiet afternoon sitting still.
- Did they go low? At least 8% of the paired readings below 4.4 mmol/L (80 mg/dL). This is the band where a wrong reading costs you the most, so it is the band the study has to actually visit, not skip past.
- Did they go high? At least 5% of the paired readings above 16.7 mmol/L (300 mg/dL). Same logic at the other end.
Only a score of 5 out of 5 tells you anything about risk. A score under 5 does not mean the device is bad. It means nobody has published enough evidence, tested in the right way, for anyone, including me, to honestly characterise how it behaves at the edges. That is not a knock on the device. It is just not something I can respond to with a straight face. So Grace will not show accuracy figures for, or recommend, any device that has not cleared all five.
Only two numbers matter after that. Not one.
Most accuracy claims lean on a single average. Averages are seductive, because they flatten a messy reality into one clean figure you can put on a poster. They are also exactly how a device with a genuinely dangerous tail can still look fine on paper, the bad readings get diluted by all the good ones sitting either side of them. I am not going to give you an average. I am going to give you two thresholds instead, because that is what actually tells you something, and only for the devices that passed Step 1.
±20/20, the no-risk band
The share of readings close enough to trust for a dosing decision without a finger prick: within 20% of your true blood glucose above 5.6 mmol/L, or within 1.1 mmol/L below it. On a device that clears the five questions, this sits somewhere in the high eighties to mid nineties.
Outside ±40/40, the high-risk band
The share of readings far enough from the truth to push a dose properly wrong, not just a bit off. On the devices that clear the five questions, this is small, usually under one and a half percent. Small is not zero.
That gap between “small” and “zero” is the whole argument for keeping a working meter within reach, on any sensor, no matter how good. Not because CGM has let you down. Because a small, honest tail exists on every device that has ever been built, and it does not announce itself. The alarm is silent. The number looks like every other number. The occasional finger prick, especially when something does not feel right, is what catches the one time in a hundred or so that the reading is lying to you.
That 1% moment only works if the backup is trustworthy too. A finger-prick meter has its own accuracy standard, and its own way of being let down by bad technique: an underfilled strip, a contaminated fingertip, a cold hand, a vial left uncapped. An inaccurate meter, or a rushed test, defeats the entire point of having a backup. Getting the technique right, clean hands, a proper drop, an in-date strip, is not a footnote. It is the other half of the safety case.
Not every finger-prick meter is equally accurate either. Every meter commissioned under NHS England Category 2 clears the ISO 15197 regulatory floor (95% of readings within 15%), but that floor hides a forty-percentage-point spread on the tighter, more clinically honest ±5% band, the one that decides whether a correction-dose decision at a low reading is safe. Ranked by that ±5% band (NHS England Category 2 commissioning list, PR00037 v5, September 2025), the top three available in the UK today:
- Contour Plus Blue (Ascensia), 86.7% within ±5%, 99.8% within ±10%, the clear standout, widest haematocrit tolerance in the list, and the only one with Second Chance Sampling
- GlucoFix Tech GK (A. Menarini), 76.2% within ±5%, also takes a blood ketone strip on the same device
- GlucoRx Smart (GlucoRx), 74.8% within ±5%, also takes a blood ketone strip on the same device
Nobody is asking you to check every reading. The other ninety-nine-point-something percent of the time, CGM is giving you a level of insight that people with T1D simply did not have a decade ago. What earns its place alongside it is the habit, and an accurate meter, not the suspicion.
Now we can talk about which one fits your life
Everything above applies before you have picked a single device. Once you know a sensor clears the five questions, and you have seen where it sits on 20/20 and 40/40, the question that is actually left is not “which is most accurate”, it is fit: which AID system it pairs with, how long it lasts on your arm, whether a nine-year-old can manage the app, whether the implant suits a sixty-year-old’s skin better than a fortnight of adhesive. That is a real question, and it deserves a real answer, just not before the one above it.
The devices that clear all five
The devices below all score 5 out of 5 on the DSN Forum UK framework (v2, March 2026), the highest tier the framework awards. Real ±20/20 and ±40/40 figures from that chart, adult data first.
Dexcom G7 and Dexcom One+
5/5. Adult ±20/20 95%, ±40/40 >99.5%, both. Same underlying hardware and accuracy data. G7 is the AID-paired route (Omnipod 5, Tandem t:slim X2 with Control-IQ, CamAPS FX); One+ is the FP10 primary-care route into the same family, not AID-paired.
FreeStyle Libre 2 Plus and 3 Plus
5/5. Adult ±20/20 94%, ±40/40 >99.5% (both). Two AID lanes inside one family: Libre 2 Plus with Omnipod 5, Libre 3 (and 3 Plus) with CamAPS FX on YpsoPump. The Libre 3 is the smallest CGM commercially available.
Roche Accu-Chek SmartGuide
5/5. Adult ±20/20 91%, ±40/40 99%. Adults 18 and over by labelling, mandatory day-one calibration, no AID partnership yet. No paediatric accuracy data published, hence no paediatric figure here.
MiniMed Simplera Sync and Instinct
5/5, both. Simplera Sync adult ±20/20 89%, ±40/40 99%. Instinct 94%, >99.5%, the FreeStyle Libre 3 Plus figures, same Abbott hardware. The 780G’s forward-path pair: Simplera Sync shipping now at 7-day wear, Instinct shipping now at up to 15-day wear. One page covers both, plus MiniMed Go Smart, the separate MDI system built around the same Instinct sensor family.
Not on this list, and why: CareSens Air and GlucoMen iCan currently score 4/5 (data-sufficiency gap, not an accuracy problem, both are being watched). Senseonics Eversense 365 is not part of the current DSN Forum 5/5 roster at all and is not listed here.
Test your knowledge
Find your best match
Pick the three things that matter most to you. The selector scores each device against your priorities and recommends your top two.
Why only these devices? Every CGM in this selector has passed the five-question clinical framework on this page. Devices without published accuracy studies in the right populations, or without a non-adjunctive indication, are not included. The selector is a starting point for a conversation with your diabetes team, not a substitute for it.
Select 3 priorities (0 of 3 chosen)