We publish our accuracy numbers.

Most calorie trackers don't tell you how accurate they are. We do.

Every nutrition app asks you to trust its numbers. Almost none of them show you how those numbers hold up against lab-verified data. We think that's backwards. If you're making daily decisions about what to eat based on an app, you deserve to know exactly how reliable it is.

This page is our answer. No cherry-picking, no marketing spin — just the numbers.

You control the precision

The single most effective way to get accurate tracking is to describe your food with detail. Clawrie's prompt mode lets you be as precise as you want — and the more specific you are, the better the result:

Vague

"milk"

↓ more specific

Better

"a glass of milk"

↓ more specific

Precise

"200ml skimmed 1.5% milk"

This is the most accurate path in Clawrie. When you describe exactly what you ate — the quantity, the brand, the preparation — the AI doesn't need to guess. It calculates from what you told it. You decide how much effort each meal is worth.

But what about photos? That's where estimation comes in, and estimation should be verified. So we tested it.

Tested in our actual pipeline

We benchmarked the AI model we ship — GPT-5.4-mini — against 743 real dishes from Google's Nutrition5k dataset. Every ingredient was weighed on a precision scale and nutritional values calculated from USDA databases. This is the gold standard for food recognition research.

Each photo was processed through the exact same pipeline that runs when you snap a meal in the app. No special tuning, no manual corrections — just the production code, cold start, no user context.

152

Calorie MAE (kcal)

33.5%

Within ±20%

743

Dishes Tested

Calorie accuracy

Accuracy bands

What percentage of calorie estimates fall within each error margin:

±10%

19.0%

±20%

33.5%

±50%

69.0%

Predicted vs actual

Each dot is one dish. The diagonal line shows perfect prediction. Points above the line are overestimates; below are underestimates. R² measures how well predictions track with actual values (1.0 = perfect).

Per-macro results

Macro	MAE	MAPE	R²	Bias
Calories	151.9 kcal	43.4%	0.326	+13.0 kcal
Protein	14.2 g	47.1%	0.302	-9.5 g
Fat	10.1 g	82.0%	0.221	-1.9 g
Carbs	17.2 g	97.2%	-0.869	+13.8 g

MAE = Mean Absolute Error (average prediction error). MAPE = Mean Absolute Percentage Error. R² = coefficient of determination (1.0 = perfect correlation). Bias = average signed error (positive = overestimate).

Error vs meal size

Bland-Altman plot showing how prediction error relates to meal size. The solid line shows mean bias; dashed lines show 95% limits of agreement.

Accuracy by calorie range

Calorie Range	Dishes	MAE	MAPE	Bias
50–200 kcal	184	70 kcal	60.5%	+40 kcal
200–400 kcal	183	146 kcal	50.9%	+110 kcal
400–600 kcal sweet spot	183	157 kcal	32.9%	+43 kcal
600–900 kcal sweet spot	158	202 kcal	29.0%	-85 kcal
900–1,500 kcal	35	361 kcal	34.7%	-353 kcal

How you can improve these numbers

The benchmark above represents a cold start — just a photo, no additional information. In real usage, you have two levers that make results better:

Add detail after the snap. Telling Clawrie "this was a large portion" or "the protein was chicken breast" gives the AI the context it needs to refine its estimate. The more specific you are, the closer it gets.
Open Food Facts grounding. After the initial estimate, Clawrie cross-references each ingredient against Open Food Facts — a verified nutritional database with over 3 million products. When a match is found, the AI recalibrates its estimate using real nutritional data instead of relying on its training alone. This grounding step is not included in the benchmark numbers above, but it systematically improves accuracy in daily use.

What to know

AI-powered nutrition estimation from photos has inherent limitations. We want to be upfront:

Portion estimation is the hardest problem. The AI identifies foods well but estimating how much is on the plate from a single photo is fundamentally challenging.
Large meals tend to be underestimated, small portions overestimated — a common pattern in vision-based estimation.
Overhead photos lose depth. A tall bowl of rice looks the same as a shallow plate from above. From experience, photos taken at a slight angle yield significantly better results — the AI can see portion height, layering, and volume. The benchmark dataset only provides overhead shots, so the numbers above reflect the harder case. In daily use with angled photos, expect better accuracy.
Mixed dishes are harder than single items. A plate with 5 components has 5x the estimation uncertainty.

Clawrie is a tracking tool, not a medical device. For most people, consistent relative accuracy — tracking trends over days and weeks — matters more than nailing any single meal. And when precision matters, that's what prompt mode is for.

Dataset: Google Nutrition5k (Creative Commons 4.0). Thames et al., "Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food", CVPR 2021. 743 dishes selected in the 50–1,500 kcal range, stratified by calorie quartile.

Last updated: March 2026

← Back to Clawrie