Can ChatGPT Count Calories? What the 2025–2026 Studies Found

---

title: "Can ChatGPT Count Calories? What the 2025–2026 Studies Found" description: "Yes, roughly. Published tests put ChatGPT, Claude and Gemini at about 14% to 36% energy error depending on what you give them. Here is when a chatbot is good enough, when a calorie tracker is better, and a prompt that gets better estimates." publishedAt: "2026-09-28" updatedAt: "2026-09-28" category: "Accuracy" tags: ["can ChatGPT count calories", "ChatGPT calorie counter", "is ChatGPT accurate for calories", "ChatGPT calorie tracker", "Gemini calorie counting", "Claude calorie counting", "AI calorie estimate accuracy"] keyTakeaways:

• "ChatGPT can estimate calories. From a photo alone, published tests put general chatbots at roughly 30% to 36% average error on energy. Given a written list of ingredients and amounts, error drops to about 14% to 18%."
• "Every photo study we found reports the same pattern: chatbots identify foods well but underestimate medium and large portions, so the error leans low."
• "The same photo can get a different answer each time. An April 2026 test of 26,904 repeated queries found median run-to-run variation from 2.4% (Claude Sonnet 4.6) to 11.0% (Gemini 2.5 Pro)."
• "The bigger limit is not accuracy but bookkeeping. A chatbot does not keep a dated food diary, running daily totals, a calorie target that adapts to your weight, or a barcode database."
• "Use a chatbot for occasional one-off estimates. Use a calorie tracker if you log every day and want to see trends. If you do use ChatGPT, give it weights, cooking fats and a fixed output format." faq:
• question: "Can ChatGPT count calories?" answer: "Yes, as an estimate. ChatGPT, Claude and Gemini can all read a meal description or a photo and return calories and macros. In a 2025 study of 195 dishes, ChatGPT-5 averaged 30.5% energy error from the photo alone and 13.9% when it was also given the ingredients. It works best for occasional estimates, not as a daily log, because it does not keep a food diary or running totals."
• question: "Is ChatGPT accurate for calorie counting?" answer: "It is roughly as accurate as a casual person guessing, and much better when you give it amounts. A 2025 validation study of 52 standardized food photos found ChatGPT-4o and Claude 3.5 Sonnet both at 35.8% mean absolute percentage error for energy, while Gemini 1.5 Pro was at 64.2%. All three underestimated larger portions. Text descriptions with weights do noticeably better."
• question: "Can I use ChatGPT as a calorie tracker?" answer: "You can, but you will be doing the tracking yourself. A chatbot does not store a structured diary, add up a day reliably across a long conversation, chart your weight, adjust your target, or scan barcodes. People who log once in a while do fine with it. People who log every meal for weeks usually move to a dedicated tracker."
• question: "Is Gemini or Claude better than ChatGPT for calories?" answer: "It depends on the model version. In the 2025 photo study, ChatGPT-4o and Claude 3.5 Sonnet tied on energy error and Gemini 1.5 Pro did worse. In a July 2026 text-based study in The Journal of Nutrition, all four models tested (ChatGPT 5.2, Claude Opus 4.5, Gemini 3 Pro Preview and Llama 4 Maverick) agreed well with reference values for energy and macros, and the differences showed up in micronutrients."
• question: "Why does ChatGPT give different calorie numbers for the same meal?" answer: "Chatbots generate each answer fresh, so the portion guess can shift between runs. In an April 2026 test that repeated the same 13 food photos 26,904 times, median run-to-run variation ranged from 2.4% to 11.0% depending on the model. Giving exact weights and asking for a fixed table format reduces the swing."
• question: "Is a calorie tracking app more accurate than ChatGPT?" answer: "Not automatically. AI tracking apps use similar vision models and also run low on photos alone. The difference is in the workflow: a tracker with a barcode scanner gives exact label values for packaged food, a review screen lets you fix the portion before saving, and the diary keeps your totals and trends. In our own 20-meal test Callie averaged 13% mean absolute error, which is our test and not an independent one."

---

Callie is our app, and it is one of the tools this article compares ChatGPT with. We have tried to answer the question on the evidence first. The studies cited below are independent of us, and the full list with dates is at the bottom.

The 30-second answer

Yes, ChatGPT can count calories, roughly. Give it a photo alone and published tests put general chatbots at about 30% to 36% average error on energy, usually on the low side. Give it the ingredients and amounts in writing and the error falls to about 14% to 18%. That is fine for an occasional "how much was that?" check. It is not a calorie tracker, because it does not keep a food diary, running totals, a target that adapts to your weight, or a barcode database. If you log every day, a dedicated tracker does that bookkeeping for you.

The same answer applies to Gemini and Claude, with differences between model versions that we cover below.

What the studies found

We found five tests from 2025 and 2026 that compared chatbot calorie estimates against known reference values. They used different models and methods, so the numbers are not directly comparable, but they point the same way.

From a photo alone: about a third off

• ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro, 52 standardized food photos. In a 2025 validation study in Current Developments in Nutrition, ChatGPT-4o and Claude 3.5 Sonnet both had 35.8% mean absolute percentage error for energy. Gemini 1.5 Pro was at 64.2%. All three models underestimated more as portions got bigger.
• ChatGPT-5, 195 dishes. A November 2025 study in Nutrients tested ChatGPT-5 on dishes from Allrecipes, the SNAPMe dataset and home-prepared meals. From the image alone, energy error averaged 123 kcal, or 30.5%.
• ChatGPT-4, 114 meal photos. A February 2025 study in Nutrients used 38 meals photographed at three portion sizes. ChatGPT identified the foods with 93.0% precision and handled small portions well, but it underestimated the weight of medium and large portions, which pulled its nutrient estimates down.

The pattern is the same one dedicated photo apps show. An NIH analysis of four popular photo-based calorie apps found they underestimated meals by roughly 250 to 345 calories when nobody corrected the portions. Recognizing food is the easy part. Judging how much of it there is, and seeing the oil and butter, is the hard part for any camera.

With a written description: much closer

The same ChatGPT-5 study also ran each dish with more context. Energy error fell from 30.5% (photo only) to 24.4% with basic descriptors, 13.9% with the photo plus detailed ingredients, and 18.1% with the ingredients alone and no photo. In other words, what you type matters more than which model you use.

A text-only study in The Journal of Nutrition (published July 2026) entered commonly eaten US foods as text into ChatGPT 5.2, Claude Opus 4.5, Gemini 3 Pro Preview and Llama 4 Maverick and compared the answers with a research food database. Agreement for energy and macronutrients was high for all four. The weak spot was micronutrients such as vitamin D, folate and iron, where every model except Claude Opus 4.5 had poor agreement on at least one nutrient.

The same meal, a different answer

Chatbots write each answer fresh, so the portion guess moves between runs. In April 2026, the diabetes technology site Diabettech sent 13 food photos to four models 26,904 times and measured carbohydrate estimates. Median run-to-run variation was 2.4% for Claude Sonnet 4.6, 8.4% for GPT-5.4, 10.3% for Gemini 3.1 Pro Preview and 11.0% for Gemini 2.5 Pro. The worst single case, a paella photo on Gemini 2.5 Pro, ranged from 55 g to 484 g of carbs. That was a blog test, not a peer-reviewed study, but the method is transparent and the size of the sample is large.

For calorie counting, a 10% swing on a 600 kcal dinner is 60 kcal. Over a day that is noise. Over a single outlier it can be a whole snack.

Where ChatGPT falls short as a calorie tracker

Accuracy is only half the job. The other half is keeping the record, and that is where a general chatbot is weakest.

1. No food diary. ChatGPT's memory can remember that you are trying to lose weight, but it does not keep a dated, editable list of what you ate each day. Chat history is not a log you can search, correct or export.
2. No reliable running totals. Adding up a day across a long conversation is exactly the kind of bookkeeping chatbots get wrong or quietly restart. You end up checking the math yourself.
3. No targets or trends. There is no daily calorie target that re-tunes from your weight trend, no weekly average and no weight graph unless you build one.
4. Portion guessing from photos. As the studies show, the model estimates portion size from the image and tends to go low on bigger plates. Nothing forces you to check that guess before you accept it.
5. Inconsistency between runs. Ask twice and you may get two numbers.
6. No barcode database. For packaged food the label is the right answer. A chatbot reading a photo of a wrapper is estimating something a barcode scan would give you exactly.
7. No health sync. It does not read your weight from a smart scale or write calories to Apple Health or Health Connect.

When ChatGPT is fine, and when a tracker is better

Feature	Callie	ChatGPT / Gemini / Claude	Manual tracker (e.g. Cronometer free)
Estimates calories from a photo			Gold only
Estimates from a typed or spoken description
You review the estimate before it is saved		Up to you
Dated food diary
Running daily totals and macros		Manual
Adaptive calorie target	Every 2 weeks
Barcode scanner
Apple Health / Health Connect sync
Streaks and weight trend			Weight trend
Answers any nutrition question	AI coach

A chatbot is fine when:

• You want a one-off estimate: a restaurant meal, a recipe you are deciding whether to cook, a sanity check on a label.
• You already know roughly what you eat and only check now and then.
• You want an explanation, for example why your stir-fry is higher than you thought, or how to hit 120 g of protein.

A tracker is better when:

• You log every day and want totals and a weekly trend without doing math.
• Your goal depends on a target that should adjust as your weight changes.
• You eat a lot of packaged food and want exact barcode values.
• You want your data in Apple Health or Health Connect with your weight and activity.

Where Callie fits

Callie is an AI calorie tracker, so it uses the same kind of model a chatbot does to estimate a meal. The difference is what happens around the estimate.

• You confirm every photo estimate. A scan opens a review screen with the detected foods, portions, calories and macros. Nothing is logged until you edit what is wrong and confirm. That is the step that fixes the low-portion bias the studies found.
• Depth-aware portions. Callie reads the depth and camera metadata your phone stores in the original photo to estimate size. Screenshots lose that data.
• A diary, targets and streaks. Every log goes into a dated diary with daily totals. Your calorie target re-tunes every two weeks from your weight trend, and a GitHub-style streak dashboard shows which days you logged. Cheat days you schedule in advance don't break the streak.
• Barcode and menu scanning, plus voice and text logging, and sync with Apple Health and Health Connect.

In our own 20-meal kitchen-scale test Callie averaged 13% mean absolute error. That is our test, not an independent one, and the method is on the Callie vs Cal AI page. Callie is not free: you get 3 AI logs to try it, then it is $34.99 a year in the US. It has no web app and fewer reviews than older trackers. If you only want the occasional estimate, a chatbot you already use costs you nothing extra.

A prompt to use if you do use ChatGPT

If you want to use ChatGPT, Gemini or Claude for calories, this prompt fixes most of the problems above. It asks for amounts you can check, flags hidden fats, and keeps the output in the same format every time so you can add it up.

You are estimating calories for a food log. For the meal below:
1. List each food as a separate row with the amount in grams or ml.
   If I gave an amount, use mine. If you had to guess, mark it "(guess)".
2. Include cooking oil, butter, sauces and dressings as their own rows.
   If I didn't mention them, ask me before assuming none were used.
3. For each row give kcal, protein, carbs and fat, using USDA values
   for plain foods and the brand's label for branded foods.
4. End with a total row. Do not round the total to a friendly number.
5. Say which row you are least sure about and how much it could change
   the total.

Meal: [describe it, e.g. "chicken thigh 180 g cooked, white rice 1 cup
cooked, broccoli, cooked in olive oil, restaurant portion"]

A few habits help more than any model choice:

• Weigh or measure the calorie-dense parts. Rice, pasta, oil, nuts, cheese and meat drive most of the error. A photo guess on these is where the 30% comes from.
• Name the cooking method and fat. "Pan-fried in a tablespoon of oil" adds around 120 kcal the camera cannot see.
• Use the label for packaged food. Type the numbers from the label instead of asking for an estimate.
• Keep the running total yourself, in a note or spreadsheet, rather than asking the chatbot to remember the day.

For the photo side, our guide to tracking calories from a photo covers angles, size references and the foods that fool every model.

When ChatGPT is the better choice

Stick with a chatbot if you only need an estimate a few times a week, you don't want another app on your phone, or you care more about understanding your food than logging it. General chatbots answer follow-up questions better than any tracker's built-in coach, and for text descriptions with amounts their energy estimates now agree well with research databases. If what you want is an answer rather than a record, you don't need a tracker.

Not sure which tool fits your habits? The which calorie tracker guide gives one pick for each kind of user.

Related reading

• AI calorie app accuracy study 2026, on why photo estimates run low.
• How to track calories from a photo, step by step.
• How to track calories accurately, for the habits that matter more than the app.
• Callie vs Cal AI, with our 20-meal benchmark.

Sources

1. Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S (2025). "Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images." Current Developments in Nutrition. https://pmc.ncbi.nlm.nih.gov/articles/PMC12513282/
2. Rodríguez-Jiménez et al. (November 19, 2025). "Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios." Nutrients 17(22):3613. https://pmc.ncbi.nlm.nih.gov/articles/PMC12655113/
3. O'Hara et al. (February 7, 2025). "An Evaluation of ChatGPT for Nutrient Content Estimation from Meal Photographs." Nutrients. https://pmc.ncbi.nlm.nih.gov/articles/PMC11858203/
4. Lawabni R, et al. (July 2026). "Evaluation of Energy and Nutrient Estimates from Large Language Models Using Text-Based Queries." The Journal of Nutrition. https://pubmed.ncbi.nlm.nih.gov/42401322/
5. Diabettech (April 15, 2026). "I Asked AI to Count My Carbs 27,000 Times. It Couldn't Give Me the Same Answer Twice." https://www.diabettech.com/i-asked-ai-to-count-my-carbs-27000-times-it-couldnt-give-me-the-same-answer-twice/
6. Charles O, Hengist A et al. (July 2026). "Accuracy of photo-based dietary tracking applications." Presented at NUTRITION 2026. Press release: https://www.eurekalert.org/news-releases/1136415
7. Cronometer Support. "Apple Health & Apple Watch." https://support.cronometer.com/hc/en-us/articles/360020734212-Apple-Health-Apple-Watch
8. Cronometer press release (September 8, 2025). "Cronometer Launches Premium Photo Logging." https://www.prnewswire.com/news-releases/cronometer-launches-premium-photo-logging-fast-verified-nutrition-tracking-for-real-life-302549752.html
9. Callie internal 20-meal kitchen-scale benchmark (May 2026). Methodology at https://www.mycallie.app/en/compare/callie-vs-cal-ai