Which AI food scanner app is the most accurate?

Because accuracy is a measurement, and a measurement needs three things: a fixed set of test meals, a ground truth (label, database, scale) and a rule for what counts as a hit. The moment any of those differs between two apps, the numbers stop being comparable. A cheeseburger with published chain nutrition is easy; a plate of curry from an independent restaurant with no reference is hard. Count the easy cases and you get 98%; count the hard ones and you get 80%.

Web listicles that assign "accuracy scores" to ten apps almost never say which meals were tested or what they were compared against. We read them the way you should: as opinion. A defensible comparison would photograph the same plates in every app and score them against weighed references. We are preparing one for German dishes, but it is not published, so it does not appear here as a result.

On 16 June 2026 we ran 16 meals from four cuisines (German, US, Japanese, Italian) through the same analysis path the app uses for a real photo, no test mode. Every result was scored deterministically against USDA FoodData Central, Open Food Facts or published restaurant nutrition; a guard verifies that every source the analysis cites exists and supports the number.

Result: 16 of 16 cases inside the ±20% per-100g gate, ~96% weighted-macro accuracy across the set. On branded items with a label, the mean error was about 3.5% (median ~2.5%): Kinder Joy 0.0%, Nutella 0.2%, Skyr 1.5%, Pocari Sweat 4.4%, with Chobani the outlier at 11.7%. Loose whole foods were harder: apple 9.9%, chicken breast 10.1%, banana 12.4%. Every number, the method and the limits are on the accuracy page.

SnapCalorie's App Store listing says its measurement is "backed by peer-reviewed research on thousands of measured dishes" and uses the iPhone's depth sensors for portions. That is the most concrete claim in the field, though the listing gives no error figures. Yazio's own help centre describes its photo results as an approximate estimate. MyFitnessPal states no test for "Meal Scan"; neither does Cal AI, and App Store reviewers report confusions such as every white meat being read as grilled chicken.

To be fair: Cronometer has the cleanest per-nutrient data of anyone here, with up to 95 lab-analysed values. If micronutrients are what you care about, it beats any photo app on that axis. And MyFitnessPal's database contains nearly every packaged product you will meet. Photo accuracy and database breadth are two different things.

Take three labelled items (a yoghurt, a granola bar, a ready meal), weigh one serving of each and photograph it. Compare the app's number with label × weight. Then one loose food, say 150 g of cooked rice or a weighed banana, and see how far the portion estimate drifts. That is roughly the logic of our benchmark, applied in your kitchen.

Watch whether the app lets you correct the result and whether it says where the number came from. A number without a source is a guess; a number tied to a label is a measurement. Our food pages give per-100g reference values with sources you can check against.

For losing, maintaining and a daily overview, an estimate with 5–15% scatter is enough, because errors average out over days and the trend is what matters. Kairo additionally learns your real energy expenditure from your weight trend, which cancels systematic estimation errors over weeks. For contest prep, medical prescriptions or gram-exact macros, a scale plus manual entry beats any AI, and that holds for every app on this page.

If you want to hold our numbers against your own plate: three AI analyses per account work in Kairo without a subscription, and barcode lookups are unlimited. After that Kairo costs $3.99 per week or $44.99 per year (3,99 € / 49,99 € in Germany), with a 3-day free trial on the yearly plan only. Details on the pricing page. Price is not an argument for Kairo: ad-funded apps with a permanent free tier are cheaper to start with.