Can ChatGPT Count Calories? What 4 Studies Show (2026)

ChatGPT can count calories. You send it a photo or describe what you ate, and it'll give you a number. Add the ingredients and amounts, and you can get a reasonably useful estimate.

But how much can you trust it when you're trying to lose fat, build muscle, or follow an actual nutrition plan?

That's where the details matter. In two photo tests, ChatGPT's average calorie error was around 30 to 36%. Giving it the ingredients and their amounts brought that down to about 14% in one of those tests. Better, definitely. But you still need to know what you're asking it to work out for you.

Let's go through what it gets right, where the mistakes come from, and how to make this easier if you're logging food every day.

How accurate is ChatGPT calorie counting?

If you tell ChatGPT exactly what you ate, including the amounts, it has a lot less to figure out. If you just send a photo of dinner, you're asking it to identify every ingredient AND work out how much of each one is there.

That's a much harder job, especially with mixed meals, sauces, or anything sitting underneath something else on the plate.

In a direct comparison, ChatGPT and Claude both had an average calorie error of 35.8%, while Gemini's was 64.2%. Another test found that ChatGPT-5's error dropped from 30.5% to 13.9% when it got a full ingredient list with amounts.

So yes, giving it more information makes a difference. Here's what the four papers tested:

Study	What was tested	Result
Three-model photo test, 2025	ChatGPT, Claude, and Gemini estimated nutrition from 52 photos of weighed food.	Average calorie error was 35.8% for both ChatGPT and Claude, compared with 64.2% for Gemini. Larger portions tended to produce larger underestimates.
Portion accuracy test, 2025	ChatGPT analyzed 114 photos showing 38 meals in three portion sizes.	Food identification precision was 93%, but meal weight was underestimated in 76% of photos. Weight estimates were better for small portions than for medium and large ones.
Four-model carb count, 2026	Four models estimated carbs from 124 written meal descriptions, compared with clinicians' calculations.	ChatGPT had the smallest bias and was the only model to meet the study's 5% equivalence criterion. That does not mean every meal was within 5%.
Added-context test, 2025	ChatGPT-5 estimated calories for 195 dishes, with increasing detail supplied alongside the images.	Average calorie error fell from 30.5% with a photo alone to 13.9% with a full ingredient list and amounts.

These tests used different meals, model versions, and methods. Sources checked September 2026. Tell us if we've misread a result.

These are averages across the meals tested. Your lunch could come out much closer, or much further off. You can't take 30% off every answer or add it back and expect the count to be fixed.

Can ChatGPT count calories from a photo?

Yes. And recognizing the food is something it can do pretty well. In a test of 114 meal photos, ChatGPT identified foods with 93% precision, but underestimated meal weight in 76% of the photos.

Think about what that means when you use it. You upload your chicken and rice, it names both correctly, and the answer looks good. But did it get the AMOUNTS right? How much rice is underneath the chicken? Did it account for the oil you cooked with?

You know some of that because you made the meal. ChatGPT only has what you gave it, and the photo may leave out quite a bit.

This is why checking the ingredient names alone isn't enough. You need to look at the portions too. In that same test, small portions were easier to estimate than medium and large ones.

Why this matters when you're trying to lose fat

Let's say your dinner has 700 calories, but the estimate comes back 35% low. You log 455 calories. You actually ate 700.

Those 245 calories never made it into your log. If you're deciding what else to eat based on that total, you're working from the wrong number before you've even finished the day.

One meal like that doesn't tell you whether your diet will work. But if you keep accepting portions that are smaller than what you served, your recorded intake will keep coming out lower than your actual intake. That's a frustrating thing to discover after you've spent weeks trying to be consistent.

How to make ChatGPT more accurate at counting calories

The most useful thing you can do is describe what the photo doesn't tell it. You don't need an elaborate prompt. You need the food, the amounts, and the details that affect the count.

1. Give it the ingredients and amounts

If you weighed the chicken and rice, include those weights. Say whether you weighed them raw or cooked. If you're using a recipe, give it the ingredients and explain how much of the finished meal you ate.

The test that got ChatGPT-5's error down to about 14% supplied a full ingredient list with amounts. It wasn't just a photo with "please be more accurate" underneath.

Of course, this takes some effort. But if you already know the amounts, there's little reason to make the model guess them again.

2. Include the cooking details

Oil in the pan, butter on the vegetables, dressing mixed through a salad... you may barely see any of it in the photo.

Tell ChatGPT what you added and how much. Otherwise, even a decent estimate of the main ingredients can leave part of the meal out. The same goes for a sauce or filling it hasn't identified.

3. Check what it counted before you use the total

Ask for the ingredients and portions separately. Then read them. If the answer lists a small serving and you know you had a large one, correct it before adding the calories to your day.

That's the part you still have to do yourself. A detailed answer can look convincing, but you know more about what you ate than the photo does.

And if you're already checking ingredients, entering weights, and doing this meal after meal, a food tracker starts to make a lot of sense. You can save the meal and use it again, with the amounts already there. We go further into that in ChatGPT vs a fitness app.

Can Claude count calories?

Yes, much like ChatGPT. If you already use Claude, you can describe your meals or send it photos and ask for a calorie estimate.

But switching to Claude just for accuracy? The photo comparison doesn't give you much reason to. Both models had the same average calorie error of 35.8%, and their portion weight errors were close too: 37.3% for Claude and 36.3% for ChatGPT.

Claude also performed worse in the separate test of written carb descriptions. It still needs the same help from you with ingredients and amounts.

ChatGPT vs Claude vs Gemini for calorie counting

ChatGPT and Claude came out about the same on the photo test. Gemini had larger errors. On the written carb test, ChatGPT did better than the other models.

Measure	ChatGPT	Claude	Gemini
Average calorie error from photos	The average error was 35.8%.	The average error was 35.8%.	The average error was 64.2%.
Average portion weight error	The average error was 36.3%.	The average error was 37.3%.	The average error was about 65%.
Larger portions	Underestimates grew with portion size.	Underestimates grew with portion size.	Underestimates grew with portion size.
Agreement with clinicians on typed carb counts	It met the study's 5% equivalence criterion.	It met neither the 5% nor the 10% criterion.	It met the 10% criterion, but not 5%.

Photo results come from the 2025 comparison. Carb results come from a separate 2026 test that also included DeepSeek. These are results for the versions tested, not a ranking of today's models.

With 124 written meal descriptions, ChatGPT had the smallest bias against clinicians' carb calculations and was the only model to pass the study's 5% equivalence test. Gemini and DeepSeek passed at 10%. Claude passed neither.

That 5% is a result across the test, not a promise that every meal was within 5%. Still, it's a good reason to give ChatGPT credit for what it can do with a properly described meal.

The photo comparison used 2024 model versions, and the later ChatGPT-5 test used different meals. So these results don't give us a ranking of every model available today. What they do give us is a reason to pay attention to the input, especially when the model has to estimate portions from a picture.

Can a better prompt solve the problem?

It can help. If your first message is just "how many calories?" with a photo attached, there's plenty you can add to make the answer more useful.

But think about what you're improving. Adding the recipe, portion weights, or cooking method gives ChatGPT something it didn't know. Asking it to be extra careful still leaves it looking at the same photo.

For daily tracking, you also want to know where the food values came from. Which ingredient did it use? Does that entry match the food you ate? Can you check it against the package or a food database?

If all you have is a final calorie number, you can't see which part needs correcting. That's why the breakdown is worth asking for, even when you only want a quick estimate.

How APEX makes daily calorie tracking easier

With APEX Lifting and Diet Coach, you can review the food and amounts, save the meal, and see it counted against your daily targets. And because your workouts are in the same app, you can follow your training and nutrition together.

Here's how the food calculation works:

The model identifies the food and estimates the amount. Upload a photo and it breaks the meal into ingredients, with a weight for each one.
The app calculates the calories and macros. It looks those foods up in its catalog and does the arithmetic in code. The same ingredients and gram amounts produce the same calculated total.
You review everything before saving. If it missed an ingredient or got the rice portion wrong, you can change it. Once you save, the meal goes into your daily totals and stays in your history.

The catalog has 1,429 foods. Of those, 853 trace to named food-composition records, including 733 from the USDA and 120 from national tables in other countries. Other entries include brand labels and estimates.

Now, this doesn't mean APEX can magically weigh dinner through your camera. The model can still get a food or portion wrong, and the studies above didn't test APEX. What you get is a clear way to check the estimate and keep using the corrected meal, without having to rebuild it in a chat every time.

If all you need is a starting calorie target, try our free calorie calculator. It uses published equations and gives you separate workout-day and rest-day estimates.

If you want to log meals regularly, keep your history, and see your nutrition alongside your training, that's what APEX is built for. Already using a food tracker? Our MyFitnessPal comparison goes through the differences.

Want to try it with your own meals? Download APEX and try the full experience free for a month.

Manual and saved-meal logging are free forever, with 10 free AI analyses to try photo logging. Available on the App Store and Google Play.

ChatGPT calorie counting FAQs

Why does ChatGPT underestimate portion sizes?

Because knowing what's on your plate doesn't tell it how much you served. Think about a deep bowl of rice, sauce underneath the chicken, or oil you added while cooking. The photo leaves a lot out. In one test, ChatGPT underestimated meal weight in 76% of photos, with larger portions causing more trouble.

Is Claude more accurate than ChatGPT for calories?

Not in the photo comparison. Claude and ChatGPT both had an average calorie error of 35.8%, and their portion estimates were very similar too. Claude also did worse on the separate written carb test. If you're switching just to get a more accurate food count, these results don't give you much reason to.

Which AI is the most accurate calorie counter?

ChatGPT did best on the written carb test and about the same as Claude on photos. Gemini had larger photo errors. But those results come from specific versions of each model, so they don't tell you which one will get your next meal right. Giving the model better information about what you ate is still worth doing, whichever one you use.

Can ChatGPT count carbs?

Yes, especially when you describe the food and amounts properly. ChatGPT was the only model to pass the carb study's 5% equivalence test against clinicians' calculations. That doesn't mean every meal was within 5%. And if you count carbs to manage diabetes, the authors describe it as help alongside diabetes education, not a replacement.

Is ChatGPT accurate enough to run a cut on?

We wouldn't use unchecked photos for the whole thing. Let's say a 700 calorie dinner gets counted 35% low. That's 245 calories missing from your log, even though you ate all of them. You can still use ChatGPT, but be ready to check the ingredients and portions instead of accepting whatever total comes back.

Is a calorie tracking app more accurate than ChatGPT?

A database makes the food values and calculations easier to check. You still need the right ingredient and amount, though. APEX shows you both before saving, so if the photo estimate says less rice than you actually served, you can change it. That's useful even when the model makes a mistake. The studies here didn't test APEX's accuracy.

Sources

Fridolfsson J, Sjöberg E, Thiwång M, Pettersson S. Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images. Current Developments in Nutrition, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12513282/
O'Hara C, Kent G, Flynn AC, Gibney ER, Timon CM. An Evaluation of ChatGPT for Nutrient Content Estimation from Meal Photographs. Nutrients, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC11858203/
Zagaroli L, Caione N, Zara S, Guerra F, Zugaro A, Baroni MG, Iezzi ML, Delvecchio M. Accuracy of ChatGPT, Gemini, Claude and DeepSeek in Carbohydrate Counting. Diabetes, Obesity and Metabolism, 2026. https://pmc.ncbi.nlm.nih.gov/articles/PMC13243987/
Rodríguez-Jiménez M, Martín-del-Campo-Becerra GD, Sumalla-Cano S, Crespo-Álvarez J, Elio I. Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios. Nutrients, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12655113/

APEX is not affiliated with OpenAI, Anthropic, or Google. The findings above apply to the meals, prompts, and model versions used in each paper.