TL;DR
Roughly, but not reliably enough to use as a diary on its own. Published studies find ChatGPT is good at recognizing foods and often in the right range for calories, but portions and individual nutrients can be well off, especially from a photo of a larger meal. Giving it weights and ingredients helps. For numbers you can check, look foods up in a nutrition database.
What the studies found
These studies are small and models change quickly, so read them as a guide, not a final score.
- •Meal photos: ChatGPT-4 was tested on 114 meal photos in Nutrients (2025). It identified foods with 93% precision and did well on the weight of small portions, but poorly on medium and large meals. It differed from actual values by more than 10% for 13 of 16 nutrients, and underestimated 11 of them.
- •Packaged meals: five chatbots, including ChatGPT, Claude, Grok, Gemini, and Copilot, were tested on eight convenience store meals in Nutrients (2025). Accuracy for calories, protein, fat, saturated fat, and carbs was 70 to 90%, while sodium and saturated fat were severely underestimated. The authors say dietitian oversight remains essential.
- •More context: ChatGPT-5 was tested on 195 dishes in Nutrients (2025). Calorie and macro errors shrank as it got more detail, from a photo alone, to added descriptions, to ingredient lists with amounts.
- •Single foods: one comparison of 222 foods, summarized in the packaged-meal paper above, found ChatGPT’s calories within 10% of the actual value for about 35 to 48% of them.
Why its numbers drift
- •Portions: it has to guess how much is on the plate, and that guess drives everything else. The meal photo study saw this most on larger meals.
- •What it can’t see: oil, butter, and dressing change the calories without changing the photo or your description.
- •It’s an estimate, not a lookup: unless it’s reading a source, the answer is generated, so describing the same meal twice can give two different numbers.
- •Vague descriptions: “a bowl of pasta” could reasonably be 300 calories or 900.
How to get a better estimate
- 1.Give weights or sizes where you can: “180 g cooked chicken breast” beats “some chicken.”
- 2.List the ingredients and amounts, including cooking oil.
- 3.For packaged food, give it the label instead of the product name.
- 4.Ask it to list its assumptions, then correct any that are wrong.
- 5.Ask for a range when you’re not sure of the portion, and don’t treat one number as exact.
When an estimate is fine, and when it isn’t
An estimate can be enough for a rough sense of a meal, or to compare two options. It’s a poor fit if you’re tracking a small daily deficit, because an error that repeats every day can be bigger than the deficit itself.
Be especially careful if you’re counting carbs for insulin dosing or managing sodium for a health condition. Don’t rely on a chatbot’s estimate for those; ask your care team.
Let the AI understand and a database do the numbers
A language model is good at working out what you meant by “a grande iced latte with oat milk.” The calories should come from a food record you can check.
That’s how Calorie Cub is built. Connected through MCP, the assistant looks foods up in a catalog built from USDA data and restaurants’ own published nutrition, and Calorie Cub calculates the totals. Every food is labeled official, reviewed, or estimated. You still need to get the portion right, because no database can see your plate.
How logging meals by chat works ›Which ChatGPT plans can connect ›Food diary, not medical advice
ChatGPT and Calorie Cub can help you keep track and think through options, but neither is a doctor or dietitian, and an assistant can be wrong with total confidence. If you’re pregnant, have diabetes or another condition, take weight-loss medication, or have a history of disordered eating, talk to a clinician before changing how you eat.
Questions people ask
Is ChatGPT accurate for macros?
It can be in the right range, but less dependably than for overall calories. In the packaged meal study, saturated fat and sodium were severely underestimated. Check anything that matters against a label or database.
Are photos or text descriptions more accurate?
Detail matters most. In the ChatGPT-5 study, a photo with ingredient amounts did better than a photo alone, and also better than the ingredient list without the photo.
Are Claude, Gemini, and Grok more accurate?
The packaged meal study tested all of them, and none replaced dietitian oversight. Models change often, so a ranking from one study won’t necessarily hold for today’s versions.
Can I use ChatGPT as my calorie tracker?
For a rough day, yes. For weeks of tracking, use it alongside a tracker so the record and numbers stay consistent.
Other guides
Use ChatGPT to help you diet ›ChatGPT diet plan and meal plan prompts ›Use Claude to help you diet ›Use Gemini to help you diet ›Use Grok and Grok Bot to help you diet ›AI calorie counter: photo, voice, or chat ›How logging meals by chat works ›Try Calorie Cub free
Sign in with Google, then log from the app or connect your assistant. No card needed.
How Calorie Cub works