AI Calorie Counter: 8 Options and How Far to Trust Them

In brief

Free AI calorie counters exist, but photo estimates are rough. In a 2025 Nutrients study, ChatGPT-4 underestimated the weight of 87 of 114 meal photos, and large meals came out at 529.5 g against 798.1 g actual. A review of Cal Counter calls it free to use and Cal AI offers a 3 day trial, as of November 2026.

What prompt makes an AI calorie counter give a usable estimate?

Paste this into Claude, ChatGPT, Gemini or Whizi, attach the photo, and fill the brackets.

A white plate on a black plinth holding a small pile of glass fruit, with a brass kitchen scale beside it, in a charcoal studio under white rim light.

Estimate the calories in this meal photo. List each food you see with grams and kcal, then give a low, middle and high total. Name what a photo can't show, such as oil, butter, sauce and dressing, and ask me about each before you finalize. My details: [cooking method], [oil or butter used], [sauce or dressing], [plate size or a known item for scale].

The range stops the model from presenting a guess as a measurement. The hidden fat line matters most, because that is where photos fail. If you have no photo, replace the first sentence with "Estimate the calories in this meal: [describe it with amounts]."

Which AI calorie counter is free?

Cal Counter, Calcounter and the chatbots have a free route, while Cal AI and MyFitnessPal put photo logging behind a trial or a paid plan. The table lists what each tool says on its own page, as read in November 2026. A cell that says not stated means the page was silent.

ToolFree useSign upPaid priceAccuracy claim
Cal Counter (reviewed by ScriptByAI)Review says "No cost to use"; no limits statedNot describedNot statedSays estimates "may not be perfectly accurate"
CalcounterPage does not say whether it is freeNot statedNot statedClaims "More accurate", no data
Cal AI3 day free trialNot statedNot statedFAQ says "CalAI is about 80% accurate"
MyFitnessPalFree tier excludes meal scanningNot statedCA$27.99 a month or CA$114.99 a year, in Canadian dollars; no US price shown on the page readNone stated
Rex.FitNot statedNot statedNot stated"Remarkable accuracy", no data
ChatGPTFree plan lists Vision as "Limited"Not statedNot stated on the page text readNone stated
ClaudeFree plan: "Limited usage every 5 hours and every week"Not statedPro $20 a month, or $17 a month billed as $200 a yearImage upload not listed on the page
Whizi10 messages a day on Whizi AI, Auto and Claude Haiku, plus 5 photos a day, no cardAccount neededPro $29.99 a monthNone, see the study below

Whizi loses on one count in this table: Cal Counter's review says "No cost to use", while Whizi needs an account. Whizi works as a chat: you send a photo or a description and read the estimate back.

Notice the last column. Not one vendor page gives a test, a sample size or an error margin. The 80% line is Cal AI's own FAQ, so treat it as that company's claim. Sources are linked at the bottom: Cal AI, MyFitnessPal, Claude pricing and ChatGPT pricing.

Are AI calorie counters accurate?

They are decent on what is on the plate and weak on how much of it there is. The clearest published test is O'Hara and colleagues in Nutrients (2025), which gave ChatGPT-4 114 photos, each of 38 meals shot at a small, medium and large portion.

ChatGPT-4 underestimated large meals: 529.5 g against 798.1 g actual
Medium meals425.8 g of 580.5 g73% of trueweightLarge meals529.5 g of 798.1 g66% of trueweightone square is about 2 percent of the true meal weight

Portion size drove the error. Weight agreed well for small meals, but medium and large meals were underestimated, and 87 of 114 photos came out light. Food identification was better: the model found 463 of 547 foods, with 35 false positives and 84 misses. The authors say foods hidden inside soups, sandwiches and salads were harder to identify.

Calories looked better than weight. The median energy difference was 0.1%, 524.5 kcal actual against 525 kcal. A median hides the spread, though. The mean absolute difference in meal weight was 27.8%. Agreement with dietitians on energy was poor to moderate, an intraclass correlation of 0.56 (95% CI 0.38 to 0.72), and median fat ran 14.8% high, 19 g against 16.6 g. The paper tested one model on standardized photos. It does not mention oil or sauces by name, and it lists unseen ingredients as a source of error.

So a photo estimate is a starting point. With a 27.8% mean absolute difference in weight, treat any single figure as a range, and widen it for a creamy or oily dish.

What do top AI calorie counter pages leave out?

They leave out prices, limits and any evidence. The Rex.Fit roundup, which holds the top spot as of November 2026, lists five apps and gives no price or free limit for any of them. One heading says 2024, and its accuracy lines carry no data. Vidnoz's list of seven calculators calls one tool free and gives one 7 day trial, with no other price, and it ends with an unrelated face rating bonus.

Neither page carries a medical caution, so neither says that people with diabetes or a history of disordered eating should talk to a professional. Both name features to look for, such as a food database, but neither tells you to check a number against one. This page does both.

The same two pages also skip what to do when the photo is wrong. The sections below cover that: the three meals you will actually photograph, the check against USDA data and the ways estimates go wrong.

What should you look for in an AI calorie counter?

Both top pages list features to look for, and the useful ones are few. Check these seven before you pay, and ask for the evidence behind each.

  1. Image recognition: ask how food is identified and whether the vendor shows a test. The Nutrients study above found 463 of 547 foods identified, which is a published number to compare against.
  2. Ease of use: a simple photo step, so you log meals when you eat them.
  3. A food database: the vendor should name the one it uses. USDA FoodData Central is the public check.
  4. Integration: some apps connect to fitness trackers and health apps.
  5. Personalization: Vidnoz's page notes that calorie needs vary with body type and activity, so a generic target is a weak one.
  6. Extras such as meal planning, recipes and progress tracking.
  7. Cost: compare the free version with the paid plan, and read the table above, because the scanning feature is often the paid part.

Rex.Fit and Vidnoz both present these as reasons to pick an app. None of them is a measure of accuracy, which is why the check against USDA data further down matters more than any feature on this list.

What does a good answer look like for three real meals?

A good answer breaks the meal into foods, gives grams, flags what it assumed and ends with a range. These three examples are written for this page. They show what to ask and what to look for. They are not outputs from a named model.

  1. A chicken stir fry on rice. Prompt: "Estimate the calories in this photo. Chicken, mixed vegetables and white rice on a 10 inch plate. I cooked it in about [number] teaspoons of oil with [sauce]." A good answer lists the rice, chicken, vegetables, oil and sauce as separate rows, gives grams for each, adds a low, middle and high total, and says the oil is the line most likely to be wrong.
  1. A restaurant pasta in a cream sauce. Prompt: "This is a pasta dish from a restaurant. I don't know the recipe. Estimate the calories as a range and list the three assumptions that move the number most." A good answer says cream, butter and cheese are invisible in a photo, widens the range and tells you it can't know the kitchen's portion.
  1. A salad with dressing on the side. Prompt: "Salad of [ingredients]. Dressing: [type], about [number] tablespoons, poured on. Estimate the calories with and without the dressing." A good answer shows the two totals separately, because the dressing can outweigh the leaves, and asks whether the toppings are croutons, nuts or cheese.

In all three, the quality marker is the same. The model separates what it sees from what it assumes. If a reply gives you one number and no assumptions, ask again.

What are the best AI calorie counter prompts?

These twelve prompts cover photos, text logs, checks and reviews. Each one names its brackets so it works on the first try. They estimate. They don't prescribe, and none of them replaces a dietitian.

For a photo:

  1. Range estimate: "Estimate the calories in this photo as a low, middle and high total. List each food with grams and kcal and state your assumptions."
  2. Hidden fat: "Look at this photo and list every place oil, butter, cheese or sauce could be hiding. For each one, ask me a question that would pin it down."
  3. Scale check: "The fork in this photo is about [length] long. Use it to estimate the plate width and the food volume, then say how sure you are."
  4. Two portions: "Estimate this meal at the portion shown and at half that portion, as two separate totals."

For a text log:

  1. Describe and count: "I ate [foods with amounts, such as 150 g cooked rice, 2 eggs, 1 tablespoon oil]. Give kcal for each line and a total, and mark any line where you assumed a size."
  2. Restaurant guess: "I ate [dish] at [type of restaurant]. Give a typical range and what would push it to the high end."
  3. Packaged food: "The label says [calories] per [serving size] and I ate [amount]. Work out my total and show the sum."

To check the answer:

  1. USDA search list: "For each food above, give the exact name I should search in USDA FoodData Central and which data type to prefer."
  2. Compare my number: "You said [number] kcal for [food and grams]. Tell me the kcal per 100 g you used, so I can compare it with FoodData Central."
  3. Find the weak line: "Which single line of your estimate would change the total most if you were wrong, and by how much?"

To review a week:

  1. Spot the pattern: "Here are my logged meals for the week: [paste]. Which entries were estimates with wide ranges? List them so I can weigh those foods once."
  2. Question for my dietitian: "Write three questions about my food log that I should ask a registered dietitian. Don't give me targets."

How do you check an AI calorie estimate against USDA data?

Search the main foods in USDA FoodData Central and compare the per 100 gram value. The USDA describes the site as "USDA's comprehensive source of food composition data with multiple distinct data types". The Agricultural Research Service manages it and the National Agricultural Library hosts it.

Pick the data type to match the food. Foundation Foods holds analytical data on minimally processed foods and is updated twice a year. Survey Foods (FNDDS) covers foods as people report eating them in the national surveys. Branded Foods carries label data from commercial products and is updated monthly. The data are in the public domain under CC0, so you can use them without permission, and the FoodData Central site asks only that you credit it.

Then do the sum. Say the model gave 150 g of cooked rice at a certain kcal. Find rice in Foundation Foods, scale its per 100 gram value to 150 g and compare. A gap over about 15% means the model has the food or the portion wrong. Fix whichever one it is and rerun the prompt.

What mistakes ruin an AI calorie estimate?

Four mistakes cause most of the bad numbers, and each has a fast fix.

  1. Taking one number as fact. The fix: ask for a low, middle and high total and plan with the middle only if the range is tight.
  2. Leaving out the oil. The fix: tell the model how much fat you cooked with, because the photo can't.
  3. Photographing a small plate at a slant. The fix: shoot from above with a fork or hand in frame, since weight error grows with portion size in the study above.
  4. Logging the model's figure for a packaged food. The fix: read the label, or check the Branded Foods entry in FoodData Central.

A fifth is quieter. People retype a number from memory next day, and the assumption goes missing. Keep the model's assumptions beside the total.

Where does Whizi fit as an AI calorie counter?

Whizi is one app instead of a pile of single purpose tools, and it suits you if you want a second opinion on a meal. You can attach a photo to a chat message for the model to look at. A free account can send 5 photos a day, and the free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.

Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Side by side Compare, which puts two models on one prompt, is a Pro and Powerhouse feature, so you can send prompt 1 to two models and read where the ranges differ. Web search is on by default in each chat, which helps when you ask for a restaurant's published nutrition page.

An Excel file made in chat has no plan gate, so you can ask for your week of estimates as a spreadsheet. See Whizi pricing for what each plan holds.

What can an AI calorie counter not do?

It can't see oil, it can't weigh food, and it can't give medical or diet advice. Everything on this page is general information. It isn't medical advice and it isn't a diet plan.

If you have diabetes, a calorie or carbohydrate estimate that is off can change what you do about insulin or medicine, so work with your care team and use a method they accept. If you have or have had an eating disorder, counting calories can itself be harmful. Talk to a qualified professional before you use any counter, AI or not.

For other checks on AI output, read how to fact check AI answers. For the photo side, which AI can analyze images covers upload limits and what a model can't see.

Workflow checklist
  • Ask for a low, middle and high estimate, never a single number
  • Tell the model about oil, butter, sauce and dressing
  • Shoot from above with a fork or hand in frame
  • Check the two biggest foods against USDA FoodData Central
  • Weigh a meal you eat often once and reuse the corrected figure
Common questions

Frequently asked questions

Is there a free AI calorie counter?

Yes. Cal Counter's review page says "No cost to use", and ChatGPT and Claude both have free plans, as of November 2026. Cal AI offers a 3 day free trial rather than a free plan, and MyFitnessPal keeps meal scanning for Premium. Free tools give no accuracy data, so treat each number as a range.

Can ChatGPT count calories from a photo?

It can estimate them, with limits. In a 2025 Nutrients study, ChatGPT-4 had a median energy difference of 0.1% on 114 photos but underestimated meal weight in 76.3% of them. Add the cooking oil and sauces yourself, because the photo can't show them.

How accurate is an AI calorie counter?

It is accurate enough for a rough guide and not for exact tracking. The Nutrients study found weight errors that grew with portion size, and Cal AI's own FAQ claims about 80% accuracy without data. Check large items against USDA FoodData Central.

Is it safe to count calories with AI if I have diabetes or an eating disorder?

Talk to a professional first. An estimate that is off can affect medicine decisions for someone with diabetes, and counting can be harmful with an eating disorder. This page is general information and not medical advice.