Open Accuracy
How far our photo estimates miss, dish by dish.
Calorie apps rarely say how wrong they can be. We measure ours on plates that were weighed on a scale and publish the result here, with the number of plates, the date and the method. Where we have not measured, this page says so.
5
of 11 dish groups measured
20
weighed plates needed before a group shows any figure
2026-10-17
last updated
South Asian home cooking
No public weighed set covers dal, roti, sabzi and thalis cooked at home, so these stay not measured yet until people who turn on “I weighed this” in the app have weighed enough plates. Calculated recipe values never count as a measurement.
Dal and lentil dishes
Not measured yetToor dal, dal tadka, chana, rajma, sambar and other home lentil dishes, cooked at home and weighed.
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Roti, paratha and breads
Not measured yetRoti, chapati, phulka, paratha, puri and dosa made at home and weighed.
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Rice dishes
Not measured yetPlain rice, jeera rice, pulao, khichdi and biryani made at home and weighed.
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Sabzi and curries
Not measured yetDry and wet vegetable dishes, paneer, egg, chicken and fish curries made at home and weighed.
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Thalis and shared plates
Not measured yetSeveral home dishes on one plate, each weighed before serving.
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Every other cuisine
Home cooking from other cuisines
Not measured yetPasta, stir fries, stews, salads and every other home meal, weighed by people who turned on "I weighed this".
Not measured yet: 0 of 20 weighed plates, 20 more needed.
Cafeteria plates (Nutrition5k)
MeasuredWeighed plates from a US cafeteria, photographed from above by Google Research (Nutrition5k test split, CC BY 4.0). Mostly Western food, so it says nothing about home cooking.
Typical miss 101 kcal on 325 weighed plates
- Calories: half the plates were within 101 kcal; 8 in 10 misses ran from minus 140 kcal to +274 kcal.
- Protein: half within 4.7 g; 8 in 10 misses ran from minus 17 g to +15.4 g.
- The range the app showed held the weighed calories on 81% of these plates and the weighed protein on 84%.
325 weighed plates, run on 2026-10-17 with anthropic/claude-haiku-5.5. A stratified sample of the Nutrition5k test split (seed 20261017), each photo read by the app's own photo path (claude-haiku-5.5, with claude-sonnet-5.5 when the main item is under 0.35 confidence or the answer is unusable) as a first photo with no remembered corrections, and compared with the weighed plate.
Cafeteria plates under 100 g
MeasuredThe same Nutrition5k plates, by the weight on the scale. Small portions are where photo estimates miss most, in proportion.
Typical miss 42 kcal on 105 weighed plates
- Calories: half the plates were within 42 kcal; 8 in 10 misses ran from minus 42 kcal to +182 kcal.
- Protein: half within 1.7 g; 8 in 10 misses ran from minus 2.8 g to +7.2 g.
- The range the app showed held the weighed calories on 71% of these plates and the weighed protein on 86%.
105 weighed plates, run on 2026-10-17 with anthropic/claude-haiku-5.5. A stratified sample of the Nutrition5k test split (seed 20261017), each photo read by the app's own photo path (claude-haiku-5.5, with claude-sonnet-5.5 when the main item is under 0.35 confidence or the answer is unusable) as a first photo with no remembered corrections, and compared with the weighed plate.
Cafeteria plates of 100 to 199 g
MeasuredThe same Nutrition5k plates, by the weight on the scale. Small portions are where photo estimates miss most, in proportion.
Typical miss 79 kcal on 86 weighed plates
- Calories: half the plates were within 79 kcal; 8 in 10 misses ran from minus 106 kcal to +309 kcal.
- Protein: half within 4.6 g; 8 in 10 misses ran from minus 11.3 g to +18 g.
- The range the app showed held the weighed calories on 81% of these plates and the weighed protein on 85%.
86 weighed plates, run on 2026-10-17 with anthropic/claude-haiku-5.5. A stratified sample of the Nutrition5k test split (seed 20261017), each photo read by the app's own photo path (claude-haiku-5.5, with claude-sonnet-5.5 when the main item is under 0.35 confidence or the answer is unusable) as a first photo with no remembered corrections, and compared with the weighed plate.
Cafeteria plates of 200 to 349 g
MeasuredThe same Nutrition5k plates, by the weight on the scale. Small portions are where photo estimates miss most, in proportion.
Typical miss 155 kcal on 88 weighed plates
- Calories: half the plates were within 155 kcal; 8 in 10 misses ran from minus 153 kcal to +300 kcal.
- Protein: half within 8.4 g; 8 in 10 misses ran from minus 14.3 g to +15.9 g.
- The range the app showed held the weighed calories on 90% of these plates and the weighed protein on 86%.
88 weighed plates, run on 2026-10-17 with anthropic/claude-haiku-5.5. A stratified sample of the Nutrition5k test split (seed 20261017), each photo read by the app's own photo path (claude-haiku-5.5, with claude-sonnet-5.5 when the main item is under 0.35 confidence or the answer is unusable) as a first photo with no remembered corrections, and compared with the weighed plate.
Cafeteria plates of 350 g and over
MeasuredThe same Nutrition5k plates, by the weight on the scale. Small portions are where photo estimates miss most, in proportion.
Typical miss 168 kcal on 46 weighed plates
- Calories: half the plates were within 168 kcal; 8 in 10 misses ran from minus 296 kcal to +313 kcal.
- Protein: half within 18.4 g; 8 in 10 misses ran from minus 43.4 g to +11.9 g.
- The range the app showed held the weighed calories on 87% of these plates and the weighed protein on 74%.
46 weighed plates, run on 2026-10-17 with anthropic/claude-haiku-5.5. A stratified sample of the Nutrition5k test split (seed 20261017), each photo read by the app's own photo path (claude-haiku-5.5, with claude-sonnet-5.5 when the main item is under 0.35 confidence or the answer is unusable) as a first photo with no remembered corrections, and compared with the weighed plate.
How we measure
We photograph plates whose food was weighed on a scale, ask the app for its estimate the same way it answers you, and compare the estimate with the weighed calories and protein. Only weighed plates count. Our dish table's calculated recipe values are never treated as a measurement.
For each group we show the typical miss (the median of how far each estimate landed from the weighed value), the range that held 8 in 10 misses, the number of plates, the date of the run and the model that made the estimates. A group needs at least 20 weighed plates before it shows any figure.
When the model that reads photos changes, every figure measured on the old model is hidden until a new run exists.
Until a group is measured, Homekcal makes no claim about how close its estimates come for that food, and every photo estimate is shown as a range you can check. Nothing outside this page states an accuracy figure.