100 Days With an AI Metabolic Twin: What the Verge Test Shows
Victoria Song, senior reviewer at The Verge and author of the Optimizer newsletter, used Twin Health's metabolic digital twin for about 100 days after receiving it in June. Her column, 'My AI doppelganger tells me what to eat,' was published on Oct. 14, 2026, and documents sustained daily use rather than a brief demo. The Verge
Functionally, the twin is an AI-powered food database with color codes for predicted glucose response. Green is ideal. Orange is so-so. Red is not good. Predictions are built from user-logged food data and data from a continuous glucose monitor, or CGM, a small wearable sensor that tracks blood sugar through the day. The Verge describes the system as predicting how the user will react to certain foods from fed metabolic data, with possible future predictions for medicines.
Building Song's twin required an extensive input stack. It required 90 days of Dexcom G7 CGMs, a smart scale, a smart blood pressure cuff, multivitamins and magnesium glycinate, an Apple Watch, and blood analysis via Quest Diagnostics. Twin Health describes its AI digital twin as a real-time model of a person's unique metabolism that analyzes data from sensors and smart devices. Twin Health
Access is constrained. The service is not available for widespread download and is currently accessible only through corporate or private health plans. It is typically aimed at people with chronic metabolic issues. Twin Health describes its program as helping people heal their metabolisms, such as losing weight, managing diabetes, and sleeping better.
The intended daily loop combines automation with human support. Twin Health states the twin provides daily guidance on food, sleep, activity, stress, and more, and that members receive support from a personal team of providers and coaches. The company states it creates digital twins using real-time data from continuous glucose monitors, activity trackers, and body composition scales, plus smart-device data and lab results, to create a plan tailored to metabolism to help lose fat, maintain muscle, and feel more energetic every day. It positions the program as addressing the root causes of obesity, prediabetes, and type 2 diabetes.
The company attaches specific outcome figures to that approach. Twin Health reports 71% of participants lowered A1C, a standard average of blood sugar, below 6.5% without glucose-lowering medications except metformin, average weight loss of 27 lbs, 85% GLP-1 elimination, and 46% insulin elimination in 2025 group results of a Cleveland Clinic study published in NEJM Catalyst.
In my view, the technical interest for builders is less in the traffic-light display than in the shift from static nutrition rules to a per-user model that keeps state over time. It is fed by CGM time series, connected devices, labs, and manual logs. That design can capture differences in glycemic response between people that population averages miss. It also makes accuracy dependent on adherence. Gaps in logging or sensor wear narrow what the model can see.
The broader context for deployment is that payer-gated distribution follows from that dependency. Bundling CGMs, scales, cuffs, watches, labs, and coaching through corporate or private health plans concentrates cost and support where chronic metabolic risk is highest. It also limits independent validation. A 100-day test by a tech reviewer clarifies usability and burden, but it does not substitute for controlled outcome data.
In my view, the long arc is still encouraging. Testable, personalized feedback on food and glucose could give patients and clinicians a tighter control loop than periodic A1C checks and generic diet sheets. The work ahead is durability, logging burden, and generalization across metabolisms, medications, and daily routines. Those are tractable engineering and clinical design problems.
