Accuracy report

Is it accurate?
Here's the honest answer.

We benchmark Luno against USDA references and real meals every month. This page shows exactly how it performs — including where it still has work to do.

Last updated June 2026 · 5,200 meals scored
The bottom line
92%

of meals land within ±10% of their true calorie count.

Median calorie error across the full sample is 7%. Below, the per-macro and per-category breakdown — including the dishes where we miss.

By macro

Where the numbers land

Each macro is measured separately because precision differs. Fat is hardest because the same dish carries a wide grams-per-portion spread; calories are easiest because errors in macros partially cancel.

Calories
tolerance ±10%
91
%
within tolerance
Median error 7%
Protein
tolerance ±15%
87
%
within tolerance
Median error 9%
Carbs
tolerance ±15%
84
%
within tolerance
Median error 11%
Fat
tolerance ±15%
81
%
within tolerance
Median error 12%
By category

Where we excel — and where we don't

An average accuracy number across all meals hides where the model is strong and where it isn't. Here's the breakdown by what kind of food we're looking at.

Packaged foods with labels
Barcode + label OCR
97%
Single-ingredient dishes
e.g. grilled chicken, oatmeal
93%
Common restaurant dishes
Bowls, sandwiches, plates
88%
Mixed homemade meals
Multiple unknown portions
81%
Novel ethnic dishes
Outside our reference set
76%
Methodology

How we measure

A short, honest walk-through of how the numbers above were produced.

Step 1

USDA reference comparison

Every meal is scored against USDA FoodData Central where available. For dishes outside the database, we use hand-labelled ground truth from a registered dietitian.

Step 2

Real-world sampling

We don't grade ourselves on photos from a lab. The benchmark draws a stratified random sample from production logs across cuisines, meal types, and times of day.

Step 3

Continuous re-validation

The benchmark re-runs weekly against the latest model. Numbers on this page reflect the most recent full run; the date above tells you when.

The honest part

What we're great at — and where we're not

We'd rather tell you up front than have you discover it logging your fourth meal.

What we do well
  • Packaged foods with visible nutrition labels.
  • Common single-protein dishes (grilled chicken, salmon, eggs).
  • Staple grains and legumes with standard portions.
  • Drinks with well-known nutrition profiles.
Where we still have work to do
  • Mixed bowls with multiple unknown ingredient portions.
  • Restaurant dishes outside chain menus we've ingested.
  • Novel regional cuisines outside our curated reference set.
  • Hand-prepared homemade meals where portion guesses dominate the error.
How it improves

The feedback loop

Every correction a user makes feeds back into the reference data. The next person logging that dish gets a more accurate read.

1,840
User corrections processed last month
+312
New foods added to the reference DB
+4 pts
Accuracy improvement vs same time last year

Don't take our word for it.

Log a meal in Luno. Compare it to the label, or to your usual tracker. If we're off, tell us — every correction makes the next read better.