# How accurate is AI calorie counting?

Accurate enough to beat the food log you'd keep by hand. That's not our claim; it's what the research says.

Source: https://nexus.kushalsm.com/blog/how-accurate-is-ai-calorie-counting/ · Updated 2026-10-16

## The study everyone cites

In 2025, researchers at the University of Gothenburg gave ChatGPT, Claude and Gemini 52 standardized food photos and compared their estimates with weighed reference values ([Fridolfsson et al., Current Developments in Nutrition](https://pubmed.ncbi.nlm.nih.gov/41081011/)). ChatGPT and Claude landed within about 36% on energy per photo. Their conclusion:

> ChatGPT and Claude achieved accuracy levels comparable with traditional self-reported dietary assessment methods but without associated user burden.

Read that again. A photo, with zero effort, did as well as the manual food diary.

## The manual log was never accurate

The calorie apps' database entries look precise. What you type into them isn't. In a classic New England Journal of Medicine study, dieters who kept food records under-reported what they actually ate by 47% on average ([Lichtman et al., 1992](https://pubmed.ncbi.nlm.nih.gov/1454084/)). Forgotten snacks, guessed portions, the cooking oil nobody logs. A five-decimal number from a database multiplied by a guessed serving is still a guess.

And the biggest error of all is the meal you never log because opening the app felt like work. Most people are gone from their food log within ten weeks ([Turner-McGrievy et al.](https://pmc.ncbi.nlm.nih.gov/articles/PMC6856872/)). Zero logged meals is 100% error.

## Those were last year's models

The Gothenburg study tested GPT-4o and Claude 3.5 Sonnet, both from 2024. The models in ChatGPT and Claude today are several generations newer, with much stronger vision and the ability to search the web mid-answer. Every one of those gains reaches Nexus automatically, because Nexus uses whichever model you're already talking to.

## How Nexus gets you closer to exact

The study's one consistent weakness was big portions: the models tended to under-estimate large plates. Nexus is built so you can close that gap in seconds:

- **Add one sentence.** "Large bowl", "cooked in butter", "I had seconds". The model uses it, and it's the difference that matters most.
- **Photograph the label.** For anything in a package, the nutrition panel gives the exact numbers. That's [as good as a barcode](https://nexus.kushalsm.com/blog/you-dont-need-a-barcode-scanner/), because it's the same source.
- **Photograph the bill.** At a chain restaurant, the assistant looks up the restaurant's [own published nutrition](https://nexus.kushalsm.com/blog/log-a-restaurant-meal-from-the-bill/).
- **Fix it with one tap.** Every meal on the Nexus card is editable. If you know better, change the number; your totals follow.

## What actually moves the scale

Weight change comes from the trend over weeks, and a trend survives a meal that's off by 15%. It doesn't survive a log that stops on day eight. The tracker that wins is the one that still has entries in month three. That's the whole design of Nexus: logging is one message in a chat you already have open, so you keep doing it.

[Add Nexus to ChatGPT or Claude](https://nexus.kushalsm.com/), or try the card in the [sandbox](https://nexus.kushalsm.com/try/) first.