Arcal vs Cal AI
Arcal and Cal AI use the same core idea: photograph your plate and let AI estimate the macros. The real differences are not in the pitch, they are in pricing transparency, who owns the app, and how each has handled App Store rules.
We measured it.
62 dishes with every ingredient weighed on a scale, photographed once and run through both apps. Cal AI averaged 51.2% error per meal. Arcal averaged 24.9%.
| Arcal | Cal AI | |
|---|---|---|
| Average error per meal | 24.9% | 51.2% |
| Calorie-weighted error | 27.1% | 45.7% |
| Median error | 22.4% | 42.9% |
| Aggregate bias | -18.3% | -27.9% |
| Closer to truth | 49 of 62 dishes | 13 of 62 dishes |
Error by meal size
The error is not noise. It tracks how much was on the plate, in the same direction for both apps: small plates come back over, large plates come back under.
| Meal size | Arcal | Cal AI |
|---|---|---|
| Under 300 kcal | +1.4% | +16.7% |
| 300 to 600 kcal | -4.8% | -9.7% |
| 600 to 900 kcal | -21.0% | -32.4% |
| Over 900 kcal | -34.3% | -54.1% |
Individual results
- An 879 kcal plate returned 133 kcal by Cal AI.
- A 911 kcal plate returned 205 kcal.
- A 903 kcal plate returned 211 kcal.
- A 1648 kcal plate returned 495 kcal. Arcal returned 522 on the same photo, so neither app handled it.
- Error runs the other way on small plates too: a 40 kcal dish returned 200 kcal, and a 340 kcal dish returned 1084 kcal.
- Sent the same photograph twice, Cal AI returned 255 kcal and then 206 kcal.
How this was measured
- 62 dishes from NutritionVerse-Real, a research dataset where every ingredient on the plate was individually weighed. That weighed total is the ground truth both apps are scored against.
- Identical full-resolution iPhone photographs went into both apps. Neither was given a better image, a cleaner angle or a second attempt.
- One photo per dish, automatic, with no manual corrections and no follow-up questions answered in either app.
- The prediction that Cal AI would score 38% error was written down before any scan was run, so the result could not be rationalised afterwards. The measured 51.2% was worse than predicted.
- Measured 18 and 19 September 2026.
Where this benchmark is weak
- This benchmark was run by Arcal, on its own competitor. Weigh it accordingly, and recompute it yourself from the raw file if you would rather not take our word for it.
- Cal AI has manual portion tools, including a depth scan and a correction conversation. This test was single photo and automatic, which strips both out. Anyone using those tools will see lower error than measured here.
- NutritionVerse-Real records one canonical gram value per ingredient name rather than a per-plate weighing, and some of its rows are known to be wrong. That adds noise to both arms equally.
- 62 dishes is 62 dishes. It is enough for the gap to be statistically clear at p = 0.000005 on a sign test, and not enough to describe every kind of meal.
- Arcal is at 24.9% average error. That is less wrong than the alternative measured here, and it is not the same thing as accurate.
We are running this against the top 20 trackers next.
Same weighed plates, same photographs, same method, published the same way. Leave an email and we will send the results the day they are out. It is also Arcal’s waitlist, so this is not a neutral mailing list and we are not going to pretend otherwise.
The honest take.
The two apps solve the same problem the same way, so this comparison is closer than most. Cal AI has a longer track record and a much larger user base. Arcal's differences are structural: pricing is public before you download, the app is still independent rather than owned by the category's largest incumbent, and there is no history of an App Store enforcement action over billing design.
Questions, answered.
Be first in line.
Arcal launches soon on iPhone. Waitlist members get access before everyone else.