The benchmark

How accurate is this?

The only fair question to ask an AI calorie tracker — so we are answering it in public, meal by meal, against the app most people already use.

What this is

A public accuracy benchmark. We keep a fixed set of test meals — everyday foods, packaged products, chain restaurant items, international dishes, vague portions, even typos — each with reference nutrition hand-entered from a cited source: USDA FoodData Central, the manufacturer's label, or the chain's official nutrition page.

Every published run parses those exact inputs through the same pipeline the app uses, and every result lands on this page — the meal, the reference numbers, what TypeCal estimated, and how far off it was. No cherry-picking: misses go up alongside the wins, because the misses are the roadmap.

Results publish here as the benchmark runs. Nothing is shown until it has a verified reference.

Meal	Reference	TypeCal	Result
First run is being verified now	—	—	—

Benchmark in progress

How we score

The scoring is deterministic — a formula, not a judge. A meal passes only if all four numbers — calories, protein, carbs, and fat — land within tolerance of the cited reference. The default tolerance is ±15%. International and composite dishes get ±25–30%, because there is no single correct biryani. No AI grades any result, including ours.

Reference values are re-verified against current official sources — nutrition data goes stale, and chains reformulate. When a reference changes, we update it and say so, even when the old number flattered us.

Every run also publishes its raw data: reference values with citations, tolerances, TypeCal's estimates, per-macro deltas, and pass/fail flags for every meal. If you disagree with a number on this page, you can check it yourself.

Head-to-head: TypeCal vs a leading AI food journal

A leading AI food journal publishes its own public accuracy benchmark. We took every food on it, logged each one in TypeCal, and scored our estimate against their own reference values — an apples-to-apples rematch on their turf.

The same 68 foods they test — everyday items, chain restaurant meals, international dishes, precise portions, even deliberate typos — each with a reference value hand-entered from USDA or an official source. We ran every input through the exact pipeline the app uses and scored it against their published reference, with their own per-category tolerance. Misses go up next to the wins. Across all 68 foods, TypeCal's mean absolute error against the reference was 7% to theirs at 11%, and TypeCal landed within tolerance on 61 meals to their 57. A few of their branded references are out of date (Big Mac, Subway, In-N-Out) — we used their numbers as published anyway.

Simple 7/10 passed

Single-ingredient foods with exact USDA reference values.

Homemade 10/10 passed

Multi-ingredient home cooking. References are USDA composites; wider ±30% tolerance.

Restaurant 10/10 passed

Chain items scored against officially published nutrition.

International 13/15 passed

Non-US dishes; references cross-referenced against government nutrition databases. ±25%.

Typos 9/10 passed

Intentionally misspelled inputs — the reference is for the correctly-spelled food. ±25%.

Portions 12/13 passed

Precisely weighed foods with specific measurements.

Scored the same way as our own benchmark — deterministic, calories within the category tolerance of their cited reference (±15% simple and portions, ±25–30% composite and international). We used their reference values verbatim, even where our own grounded data suggests they're stale. References transcribed from a leading AI food journal's public accuracy page.

What's next

A MyFitnessPal column, logged by hand through its standard search flow so the comparison reflects what a real user gets — that tooling is built and capture is underway. And a larger test set: more international home cooking, more brands, more of the meals people actually eat.

Why we publish this

Every AI estimate in TypeCal is a draft you can inspect and edit before it counts — that honesty is the product. Publishing our accuracy, misses included, is the same idea at a larger scale.

It also keeps us honest with ourselves: if a category of food parses badly, this page shows it, and fixing it becomes the roadmap.