Cal AI is a photo-based calorie tracker that asks a multimodal model to identify food in an image, then retrieves nutrition data from scraped public food databases through a thin retrieval-augmented layer. The pitch is simple. Point your camera at a plate, get calories and macros, move on. The execution is a long way from that pitch. The model regularly misidentifies dishes, the database lookups produce values that do not match package labels, and the apps that should make a photo-first tracker work on Apple devices are either partial or broken.
01Data breach
The product is now better understood through its security and billing record than through its food estimates. On March 9, 2026, a hacker using the alias "vibecodelegend" posted a 14.59 GB dump on BreachForums containing more than 3.2 million Cal AI user records. The reported attack vector was an unauthenticated Google Firebase backend with the subscription table readable without credentials. Account PINs were 4 digits with no rate limiting, no CAPTCHA, and no lockout, which means a brute force attempt against the 10,000 combinations finishes in minutes. The exposed fields include full names, email addresses, dates of birth, gender, height and weight, social media profiles, meal logs with timestamps, and subscription details, with at least one record belonging to a user born in 2014. Coverage from Kiteworks and Cybernews lays out the technical detail and the regulatory exposure under COPPA and GDPR. Roughly 1.2 million exposed email addresses used Apple's private relay, which means a meaningful share of those users had specifically opted into stronger privacy protections that the backend then failed to honor.
02App Store removal
The billing record landed six weeks later. In the week of April 14, 2026, Apple removed Cal AI from the App Store for multiple App Review Guideline violations. According to MacRumors, the removal was not a generic external-payments dispute. Cal AI bypassed the required in-app purchase flow with an embedded Stripe checkout, displayed weekly calculated pricing more prominently than the actual billed amount, used a free-trial toggle that did not make automatic renewal clear, and prompted users who declined the first subscription with a second different subscription flow. The app was reinstated after the developer changed the billing design and quickly returned to the top of the Health and Fitness chart. Apple framed the action as evidence that App Store enforcement against deceptive billing remains active even after the Epic Games ruling.
03Dynamic pricing
Pricing on Cal AI is dynamic. There is no fixed price list on the website or the App Store listing, and the price you see depends on your onboarding answers, your device, and the cohort the company is currently testing. Reported tiers include a $5.99 weekly plan, $29.99 and $49.99 annual plans seen on different days and flows, and a $59.99 family plan. The 3-day free trial requires a payment method up front, and after it ends the free tier collapses to three AI scans per 24 hours with macro detail and analytics behind the paywall. Three scans is not enough to log a day of eating with breakfast, lunch, dinner, and a snack, so the free tier reads as a teaser rather than a usable product.
04Photo accuracy
Accuracy on simple, photogenic items is acceptable. Independent testing puts whole foods and clearly visible portions in an 85 to 92 percent range, and barcode scans of packaged foods are essentially perfect because the data comes straight from a database lookup. The story changes on real meals. Lifehacker put a Pink Lady apple in front of the camera and Cal AI returned tikka masala. With the apple next to a kitchen scale, Cal AI estimated 80 calories against an actual value closer to 120, a 33 percent underestimate. On a plate of fried tofu, onions, cucumbers, tomatoes, feta, and chickpeas with an oil-based vinaigrette, the same review described the results as a masterclass in algorithmic overconfidence. The structural issue is that a camera cannot see cooking oil, butter, sauce volume, or how much of the plate you actually finished, and a model trained on stock food photography defaults to average portions rather than measuring what is in the frame. Across mixed dishes, curries, stir-fries, casseroles, soups, and dressed salads, the published variance lands between 25 and 50 percent, almost always toward underestimation.
05Marketing-led growth
The growth story behind the product is marketing-led, not product-led. Co-founder Zach Yadegari and his team built three layered acquisition engines. The first was systematic outreach to TikTok and Instagram creators with native-style posts, scaled through a virtual-assistant operation that ran roughly 150 active influencers at any given time. The second was performance advertising on Meta and TikTok at roughly $5 CPM and over $1 million per month in spend by January 2026. The third was an affiliate program that paid creators per install. The combination took the app to about $30 million in 2025 revenue and $5.7 million in January 2026 alone before MyFitnessPal acquired the company in December 2025 and announced the deal in March 2026. The accuracy of the underlying photo model never had to be the reason the app grew. The acquisition engines were.
06Apple ecosystem and support
User experience after install is shaped by that same priority order. The onboarding questionnaire is invasive, the paywall is the gate, and once inside, support routes through an AI bot that cannot issue refunds. Users describing cancellation and refund attempts on Reddit, TikTok, and the App Store report a loop of automated replies that ends with a recommendation to file an Apple report-a-problem ticket or a credit card chargeback. Apple Health sync is partial, with water entries flowing through and food entries not, and the Apple Watch app is widely reported as broken. Corrections to a misidentified scan do not persist, so re-scanning the same plate the next day returns a different wrong number.
07Safety implications
The safety implications matter. Systematic underestimation of 25 to 50 percent on common meals is not a neutral error. For someone running a deliberate calorie deficit it lengthens the deficit. For people with restrictive eating patterns or histories of disordered eating it reinforces under-eating with a confident-looking number on a screen. Cal AI ships with no goal-setting safeguards, no minimum calorie floor for adult users, and no friction around aggressive weight-loss targets. The 2014 birth year that surfaced in the breach indicates the user base includes minors, and there is no age-aware coaching layer to handle that.
08External references
- Kiteworks on the 3.2 million record breach
- Cybernews on the unauthenticated Firebase exposure
- MacRumors on the App Store removal for deceptive billing
- Lifehacker on AI calorie counting app accuracy
- TechCrunch on the MyFitnessPal acquisition of Cal AI
- App Store reviews for Cal AI
09Verdict
Skip Cal AI. The combination of a 3.2 million record breach driven by a misconfigured Firebase and 4-digit PINs without rate limiting, an App Store removal for deceptive billing, dynamic pricing that hides what you will actually pay, and a photo model that underestimates real meals by 25 to 50 percent makes this the wrong place to put your food data and your money in 2026.
There is a narrow case for someone who wants a casual calorie-awareness toy, eats mostly whole foods or packaged items, treats every estimate as roughly right, and is willing to pay a weekly subscription with no expectation of accuracy on cooked dishes. For anyone running a real cut, recovering from disordered eating, tracking macros for training, or simply expecting their nutrition app to protect their data, Cal AI does not fit.
Fuel is built for the dimensions where Cal AI is weakest. Food data stays on the device with on-device privacy rather than sitting in an unauthenticated cloud table. Pricing is published and stable rather than A/B tested behind an onboarding wall. The photo-first flow lets you photograph the meal or the label, describe it in voice or text, and refine it with natural-language correction until the entry matches what you actually ate, and corrections persist so the system learns. A real coach is with you through the day rather than an AI bot that exists to deflect refund requests. A live daily health score breaks down five dimensions of how the day is going, a morning recap sets the day, and an in-depth weekly review gives you an explicit action plan for the next seven days. The Apple ecosystem is treated as first class, with deep integration across iPhone, Apple Watch, Apple Health, and Shortcuts. If you want the best AI in food logging on Apple devices and you want it built by a team whose growth strategy is the product itself, Fuel is the answer.
