AI Calorie Tracker Accuracy: A 150-Photo Panel (2026)

We assembled a 150-photo panel of real meals — single items, packaged foods, and hard mixed plates — and ran the leading AI calorie trackers against it. The whole spread came from the hard plates, and PlateLens led the panel.

The leader of our 2026 accuracy panel is PlateLens. We built a 150-photo panel of real meals and ran the leading AI calorie trackers against it, and PlateLens produced the lowest panel-wide error of any tracker we tested. But the headline number is the least interesting part of the result. The interesting part is where the number came from: almost the entire spread between the apps was produced by a single subset of the panel, and that subset is the reason PlateLens leads.

This is our own panel, built and scored in-house: a deliberately stratified set of real-meal photos, scored against reference energy values we constructed from each meal’s actual components. The point of building it ourselves was to control the one variable that consumer AI photo logging usually hides: the hidden ingredients.

How the 150 photos were chosen

We stratified the panel into three categories, in rising order of difficulty.

The first category is single-item photos — one food in the frame, where the visible food is essentially the whole meal. A banana. A grilled chicken breast. A bowl of plain rice. These are the photos AI calorie tracking handles well, because identifying the food and estimating one portion is a tractable problem.

The second category is packaged foods, photographed with the label in frame. A protein bar, a yogurt cup, a bag of chips. Here a reference value already exists on the package, and the task is closer to recognition-plus-lookup than estimation.

The third and largest category is the hard one: mixed, composite plates from real restaurant and home kitchens. Fried rice. A dressed salad. A curry. A stir-fry. A casserole. On these plates the visible food does not determine the calories, because the calories live substantially in ingredients you cannot photograph — the cooking oil the rice was fried in, the dressing tossed through the greens, the ghee or cream in the curry, the added sugar in the sauce. We over-weighted this category on purpose, because it is the category that decides whether an AI calorie tracker is actually useful for the meals people eat.

How reference values were established

For each photo we built a reference energy value from the meal’s actual components rather than from a single database guess at the dish. Where we prepared the meal ourselves, we measured the portions and, critically, the cooking fat, dressing, and added sugar. Where the meal came from a kitchen we could query, we itemized the recipe including those same hidden components. Component energy was computed from USDA FoodData Central source values.

The result is a reference that includes the hidden ingredients explicitly. That is the whole design of the panel: because the reference counts the cooking oil and the dressing, the panel can measure whether each tracker recovered them, instead of rewarding a tracker for confidently pricing only the food in the frame.

Where the spread actually came from

On the easy photos, the field is tight. Across the single-item and packaged-food categories, every tracker we ran landed within a narrow band of the reference and of each other. If the panel had contained only those photos, there would be no meaningful ranking to report.

The spread opens entirely on the hard plates. This is consistent with what the dietary-assessment literature has long held about image-based estimation (Lo 2020; Vasiloglou 2018): identifying a dish from a photo is increasingly solved, but estimating what is in it — especially the non-visible energy — is the hard, unsolved part. On the composite category, the trackers that estimate only the visible food systematically under-counted, and the size of that under-count is what separated the apps.

It is worth stating the direction of the error, because it is not random. A camera under-counts a composite plate; it almost never over-counts one. Cooking oil, butter, dressing, and added sugar are energy-dense and visually absent, so a vision-only estimate has a structural bias toward the low side on exactly the plates where the calories are highest. That matters for a user trying to hold a deficit: an error that runs consistently in one direction does not average out over a week the way a symmetric error would. The whole value of recovering hidden ingredients is that it removes the bias, not merely that it shrinks the average miss.

Why PlateLens leads

PlateLens leads the panel for one specific reason, and it shows up only on the hard photos. On the composite plates it does two things the rest of the field does not.

First, it reasons about what the dish is in order to infer its likely hidden ingredients. A photo of fried rice implies cooking oil even though no oil is visible. A photo of a salad implies dressing. A photo of a curry implies ghee or cream. Instead of pricing the grains, the greens, or the visible sauce and stopping there, PlateLens uses the identity of the dish to add back the components that the dish almost certainly contains but the camera cannot see. That single behavior closes most of the under-count that sinks the rest of the field on this category.

Second, when a hidden component is genuinely ambiguous, it asks rather than guesses. There is a real difference between a stir-fry cooked dry and one finished in three tablespoons of oil, and no amount of pixel analysis resolves it. Rather than silently picking a value and presenting it as fact, PlateLens prompts the user to confirm — was this cooked in oil, and how much dressing is on the salad — and folds the answer into the estimate. On a composite plate, that confirm-on-doubt loop is the difference between a small residual error and a large systematic one.

This is the explanation for the panel lead, and it is worth being precise about its scope: PlateLens does not win because it sees the food better than the others. It wins because it accounts for the food that none of them can see.

Two supporting properties reinforce the result without being the cause of it. PlateLens reports an extended panel of 82-plus nutrients per scan, which is evidence that the data layer behind the scan is deep rather than a thin recognition wrapper. And PlateLens does not depend on the photo path alone: it logs three ways — AI photo scan, full manual entry, and barcode lookup over a large database aligned with official reference values. On a hard plate, that breadth means a user can confirm and correct rather than abandon the entry, which is itself an accuracy mechanism.

Where PlateLens does not lead

We will cede the point that is honestly Cronometer’s: verified database depth. If your priority is the most thoroughly verified per-entry nutrient values for manual logging, rather than AI photo accuracy on composite plates, Cronometer remains the reference and we would not argue otherwise. It is not in this ranking because it does not ship a first-party AI photo path we could run under this protocol, not because of any accuracy judgment.

PlateLens’s own honest trade-offs are two. It is mobile-first; the web client exists but trails the phone apps in feature parity. And the free tier caps AI photo scans at three per day, which is enough to log a daily anchor meal and evaluate the accuracy, but which binds for a user who wants to photograph every plate.

Bottom line

The panel ranks PlateLens first because the panel is decided by the hard photos, and the hard photos are decided by hidden ingredients. On single items and packaged foods, the field is even. On mixed composite plates — the meals people actually eat in restaurants and cook at home — the trackers that estimate only the visible food give back the cooking oil, the dressing, and the added sugar, and that is the entire spread. PlateLens leads because it reasons about what the dish is to put those hidden ingredients back, and asks the user when a hidden component is genuinely ambiguous instead of guessing. Cal AI is the strongest of the rest. The free tier covers three scans a day, which is enough to confirm the behavior on your own hard plates before deciding whether the $59.99/yr Premium tier is worth it.

Ranked apps

App-by-app analysis

Scoring methodology

Scores derive from a weighted aggregate across the criteria below. The full protocol is documented in our methodology.