Marble is built around an AI you can talk to. It's also built on one rule that keeps the AI honest: code computes the numbers; the AI only collects inputs and explains results. A language model is good at reading a plate and at explaining things kindly. It's bad at arithmetic it can't show, and it will happily invent a number to be helpful. So it doesn't get to.
Who does what
| Code, in the app | The AI |
|---|---|
| Your calorie target, protein, fat and fiber | Reads a meal photo and estimates what's on the plate |
| The floors, the pace limit, the 500-calorie deficit cap | Understands "a bowl of my usual oatmeal" or "half of that" |
| Your weight trend, and correcting the target from it | Explains why your target moved, in plain words |
| How much of today is left for the meal you're about to eat | Suggests what to order or cook to fit it |
| Spotting signs that food has become hard, and the helpline card | Says something kind, and nothing about numbers |
The coach can't set or change your target. When you ask it to ("make my target 1,400"), it explains that the app sets the target from your weigh-ins, and points to the Goals screen. Every equation the code uses is cited on Sources and methods.
Where the coach's food numbers come from
- Chain restaurants: the chain's own published nutrition, which US chains with 20 or more locations must provide. Marble labels a figure "published" only when it came from that table, and keeps the page it came from.
- Plain foods (a banana, a cup of rice): calories per gram from the USDA's FoodData Central, multiplied by the portion.
- Packaged food: the label, via a barcode lookup on Open Food Facts.
- A photo of your plate: an estimate. We measure how good it is on weighed meals, and publish the results, including where it's weak: photo accuracy research and the open benchmark.
The models
Marble uses Anthropic's Claude. We change a model only after re-running our checks.
Photo reader: summary
Read the photo research- Model
- Claude Sonnet 5.5, medium effort
- In use since
- 19 October 2026 (Claude Haiku 5.5 before)
- Used for
- Every meal photo, for every user. Free and Plus get the same reader; Plus pays for more reads and the coach, not a smarter model.
- Tested on
- The practice halves of public datasets of weighed meals, read side by side with the old reader (method)
- Results
- 21 to 42 calories a plate closer than Haiku on two of three sets, never worse, at about 20 times the cost per photo (October 2026)
- Known limits
- A photo can't show oil, butter or what's under the rice. Readers we've measured run high on small plates and low on big ones.
Coach: summary
See the safety results- Model
- Claude Sonnet 5.5 (medium effort) for the coach's answers; Claude Haiku 5.5 for quicker jobs such as restaurant picks and editing a meal by chat
- In use since
- 27 October 2026 (Claude Haiku 5.5 before)
- Used for
- Answering questions, logging food and weight by message, what to eat or order
- Tested on
- 169 fixed conversations and 44 red-team cases (below, and on safety)
- Results
- 156 of 169 (Haiku 147); 18 of 18 safety; red team 43 of 44 (27 October 2026)
- Known limits
- It can name a dish a restaurant doesn't serve (1 of 60 checked). Grading is by code and a separate model, not people.
What leaves your phone
There's no account. Your meals, photos, weigh-ins, workouts and conversations are stored on your iPhone. Before Marble sends anything to AI, it asks, and lists exactly what goes:
- photos you take or attach, and meal descriptions;
- your messages to the coach;
- your goal, dietary notes, recent meals and a summary of your week, and what you've asked the coach to remember;
- figures the coach asks for: totals, targets, weight trend, workouts, and Apple Health activity and sleep if you've turned them on;
- your town and nearby restaurants, when you ask where to eat.
That goes to Anthropic, which doesn't train its models on it. The coach's spoken replies go to Deepgram to become audio. Nothing is sold, and there are no ads, analytics or trackers in Marble. Say no, and you can still type meals, scan barcodes and weigh in. The details are in the privacy policy.
How we check the coach
Every prompt or model change runs against a fixed set of 169 coach conversations before it ships. They check that it uses your real numbers rather than making them up, that its food answers are right, that it stays safe, and that restaurant picks are real places with real dishes. A separate red-team set tries to talk it into harm (safety, and those results).
| Check (27 October 2026) | Result |
|---|---|
| All 169 conversations, the coach now in the app | 156 of 169 |
| Same set, the smaller model it replaced | 147 of 169 |
| Safety cases within it | 18 of 18 |
| Restaurant dishes it named, checked against real menus | 57 of 60 real, 1 invented |
The conversations are graded by code (did it call the right tool, did it use your figures) and by a separate grading model against a written rule per case. The invented dish was an açaí bowl at a Brazilian restaurant whose menu doesn't list one. We count misses like that rather than round them away.
Changes
- Page published.
- The coach's answers moved from Claude Haiku 5.5 to Claude Sonnet 5.5: 147 → 156 of 169; restaurant answers name real dishes from the menu.
- Meal photos moved from Claude Haiku 5.5 to Claude Sonnet 5.5, for every user.
Something here wrong or out of date? Write to support@getmarble.app.